text
stringlengths 0
643k
⌀ | meta
stringlengths 137
151
⌀ |
|---|---|
null | null |
# Introduction
learning techniques play a central role in data analysis, data fusion and visualization. As geophysical acquisition tools become more sophisticated and gather more information, data analysts relay more on machine learning techniques for generating meaningful representations of the data. A coherent representation of complex data often includes a feature extraction step followed by a dimensionality reduction step, which results in a compact and visual model. Analysis tasks such as clustering, classification, anomaly detection or regression may be carried out in the constructed low-dimensional space. Common dimensionality reduction methods such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) project the feature space into a low dimensional space by constructing meaningful coordinated that are linear combinations of the original feature vectors. PCA is widely used for low-dimensional modeling of geoscience datasets. Jones & Christopher applied PCA to infer aerosol specification for research of oceans or more complex land surfaces. Griparis and Faur applied a linear dimensionality reduction tool, Linear Discriminant Analysis (LDA) for a projection of earth observations into a low-dimensional space. Their low-dimensional representation resulted in a cluster organization of the image data by land types. PCA and Self organization maps were applied for pattern recognition in volcano seismic spectra by Unglert et. al. and for geologic pattern recognition by Roden et. al.. Another key issue in processing large amounts of data is the ability to fuse data from different sensors. Typical seismometers record data using three channels. These three channels capture the motion in the horizontal and perpendicular directions to the earth. Each channel may be processed separately and the results can be combined. Alternatively, a fused representation may be formed for common analysis. Recent advances in machine learning and in particular the use of non-linear kernel-based algorithm enable to construct data-driven fusions and to compute geometry-preserving low-dimensional embeddings. Such kernel-based embedding techniques are known as manifold learning methods, among them Local Linear Embedding, Lapacian Eigenmaps and Diffusion Maps (DM). Manifold learning methods overcome limitations of linear dimensionality reduction tools such as PCA and LDA. When the relationship between the original high-dimensional points is complex and non-linear, linear projections may fail to organize the data in a way that is loyal to the intrinsic physical parameters that drives the observed phenomena. This work focuses on extending manifold learning techniques for low-dimensional modeling and kernel based data-driven fusion of seismic data. Identifying the characteristic of seismic events is a challenging and important task. This includes the discrimination between earthquakes and explosions which is not only an essential component of nuclear test monitoring but it is also important for the maintaining the quality of earthquake catalogs. For example, wrong classification of explosions as earthquakes may cause the erroneous estimation of seismicity hazard. The discrimination task is typically performed based on some extracted seismic parameters. Among such parameters is the focal depth, the ratio between surface wave magnitude and body wave magnitude and the spectral ratio between different seismic phases ,. Discrimination methods based on seismic parameters give only a partial solution to the problem. For instance, a larger half of seismic events reported by the Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) are not "screened out" as natural events or even are not considered for the discrimination at all although most of those events are typically earthquakes. Recently, this problem and other geophysical challenges have been approached using machine learning frameworks. Hidden Markov model were proposed in, , and modeled the data in an unsupervised manner. Artificial neural networks, , or support vector machines were also used to construct a classifier in a supervised manner. The study in utilizes Self Organization Maps to distinguish micro-earthquakes from quarry blasts in the vicinity of Istanbul, Turkey. Manifold learning is used in for seismic phase classification. In a graph is used to detect sea mines in side-scan sonar images. The DM method is used in for visualization of meteorological data. A non-linear dimensionality reduction is proposed in to discriminate between earthquakes and explosions. In this study, the manifold learning approach that was presented in is extended by using a kernel-based fusion method for identification of seismic events. The method is model-free, and it is based on signal processing for feature extraction followed by manifold learning techniques for embedding the data. Furthermore, the method reviles the underlying intrinsic physical properties of the data, which results in a natural organization of the events by type. Since seismic data is recorded at multiple channels, we suggest fusing the information to extract a more reliable representation for the seismic recordings. The fusion framework is based on a recent work by. The study extends Diffusion Maps (DM), which has been successfully applied for phase classification, for estimation of arrival times and for events discrimination. Other constructions for fusing kernels were proposed in. The proposed framework begins with a preprocessing stage in which a time-frequency representation is extracted from each seismic event. The training phase includes the construction of a normalized graph that holds the local connections between the seismic events. A low dimensional map is then obtained by the eigen-decomposition of the graph. The constructed embedding is distance preserving. Thus the geometry of the dataset is kept in the new embedding coordinates. By utilizing the low dimensional embedding, we demonstrate capabilities of classification, location estimation and anomaly detection of seismic events. The paper is organized as follows: Sections [2](#sec:Manifold){reference-type="ref" reference="sec:Manifold"} and [3](#SecMulti){reference-type="ref" reference="SecMulti"} present the machine learning frameworks for manifold learning and data fusion. In Section [4](#sec:Data){reference-type="ref" reference="sec:Data"} the data set is described. The mathematical methods required for analysis of seismic data are provided in Section [5](#sec:Pre){reference-type="ref" reference="sec:Pre"}. The proposed framework and experimental results are presented in Section [6](#sec:Exp){reference-type="ref" reference="sec:Exp"}. We conclude this work in Section [7](#sec:Future){reference-type="ref" reference="sec:Future"}.
# Manifold Learning {#sec:Manifold}
This section reviews the manifold learning method that is applied in this work for non-linear dimensionality reduction, diffusion maps. The method's main ingredient is a kernel function. Here, radial basis kernel functions are used, their construction is described in detail.
## Radial Basis Kernel Function
Kernel functions are vastly utilized in machine learning. Classification, clustering and manifold learning use some affinity measure to learn the relations among data points. A kernel is a pre-defined similarity function designed to capture the fundamental structure of a high dimensional data set. Given a high dimensional data set \( {\myvec{X} = \lbrace { { \myvec{x}_1,\myvec{x}_2, \myvec{x}_3...,\myvec{x}_M }} \rbrace },\myvec{x}_i \in {\mathbb{R}^{D}}, \) a kernel \({{\cal{K}}: \mymat{X}\times{\mymat{X}}\longrightarrow{\mathbb{R}} }\) is an affinity function over all pairs of points in \(\mymat{X}\). The discrete kernel is represented by a matrix \(\mymat{K}\) with following properties
- Symmetry \({K_{i,j}={\cal{K}}(\myvec{x}_i,\myvec{x}_j)={\cal{K}}(\myvec{x}_j,\myvec{x}_i) } \)
- Positive semi-definiteness: \({ \myvec{v}_i^T \mymat{K} \myvec{v}_i \geq 0 }\) for all \(\myvec{v}_i \in \mathbb{R}^M\) and \({{\cal{K}}(\myvec{x}_i,\myvec{x}_j) \geq 0. } \)
These properties guarantee that the matrix \(\mymat{K}\) has real eigenvectors and non-negative real eigenvalues. In this study radial basis functions (RBF) are used for constructing the kernel. The RBF kernel function is defined by \[\label{eq:RBF} { K_{i,j}=exp\{ {-\frac{||\myvec{x}_i-\myvec{x}_j||^2}{2 \sigma^2}\}} }.\] Applying the Euclidean distance to high dimensional pairs of distant vectors could somewhat be misleading, as data is typically sparse in the high-dimensional space. For this reason the decaying property of the Gaussian kernel is beneficial. The Gaussian tends to zero for distant points, whereas its value is close to one for adjacent points.
## Setting the Kernel's Bandwidth
The kernel's bandwidth \(\sigma\) controls the number of points taken into consideration by the kernel. A simple choice for \( \sigma \) is based on the standard deviation of the data. This approach is good when the data is sampled from a uniform distribution. In this study, we use a max-min measure. The method was proposed in and aims to find a small scale to maintain local connectivities. The scale is set to \[\label{eq:MaxMin} \sigma^2_{\text{MaxMin}}={\cal{C}}\cdot \underset{j}{\max} [ \underset{i,i\neq j}{\min} (||\myvec{x}_i-\myvec{x}_j||^2)],\] where \({\cal{C}} \in [2,3]\). Alternative methods such as have demonstrated similar results in our experiments.
## Non-Linear Dimensionality Reduction {#SecDiff}
Most dimensionality reduction methods are unsupervised frameworks that seek for a low dimensional representation of complex, high dimensional data sets. Each method preserves a certain criteria while reducing the dimension of the data. Principal component analysis (PCA), reduces the dimension of the data while preserving most of the variance. Non linear methods such as Local Linear Embedding, Laplacian Eigenmaps, Diffusion Maps (DM) preserve the local structure of the high-dimensional data. In particular, in DM, a metric that describes the intrinsic connectivity between the data points is defined. This metric is preserved in the low-dimensional space, resulting in a distance-preserving embedding. The metric is refereed to as diffusion distance, it is defined later in this subsection. The DM framework enforces a fictitious random walk on the graph of a high dimensional data set \(\myvec{X}=\{\myvec{x}_1..,\myvec{x}_M \},\myvec{x}_i \in \mathbb{R}^D\). This results in a Markovian process that travels in the high-dimensional space only in areas where the sampled data exists. The method has been demonstrated useful when applied to audio signals, image editing, medical data analysis and other types of data sets. Reducing the dimension of a data set by construction of DM coordinates is performed using the following steps
1. Given a data set \(\myvec{X}\) compute an RBF kernel \(\myvec{K}\) based on Eq. [\[eq:RBF\]](#eq:RBF){reference-type="ref" reference="eq:RBF"}.
2. Normalize the kernel using \(\mymat{D}\) where \( D_{i,i}=\underset{j}{\sum}{K_{ij}} \). Construct the row stochastic matrix \(\mymat{P}\) by \[{P_{i,j}\defeq{\cal{P}}(\myvec{x}_i,\myvec{x}_j)\defeq[{{\mymat{D}}^{-1}{\mymat{K}} }}]_{i,j} \label{EquationPDM}.\]
3. Compute the spectral decomposition of the matrix \( \mymat{P} \) to obtain a sequence of eigenvalues \({\lbrace {\lambda_m}\rbrace } \) and normalized right eigenvectors \({\lbrace{{\mbox{\boldmath\){\psi}\(}}_m}\rbrace } \) that satisfy \({ {\mymat{P}} {\mbox{\boldmath\){\psi}\(}_m} =\lambda_m{\mbox{\boldmath\){\psi}\(}}_m, m=0...,M-1}\);
4. Define the \(d\)-dimensional (\(d \ll D\)) DM representation as \[\label{EQPSI}{ \myvec{\Psi}{(\myvec{x}_i)}: \myvec{x}_i \longmapsto \begin{bmatrix} { \lambda_1\psi_1(i)}, {.} {.} {.} , {\lambda_{d}\psi_{d}(i)}\\ \end{bmatrix}^T \in{\mathbb{R}^{d}} },\] where \(\psi_m(i)\) denotes the \(i^{\rm{th}}\) element of \({\mbox{\boldmath\){\psi}\(}_m}\).
The power of the DM framework stems from the Diffusion Distance (Eq. [\[EqDist\]](#EqDist){reference-type="ref" reference="EqDist"}). It was shown in that the Euclidean distance in the embedded space \(\myvec{\Psi}{(\myvec{x}_i)}\) is equal to a weighted distance between rows of the probability matrix \(\myvec{P}\). This distance is defined as the Diffusion Distance \[{ \label{EqDist} { {\cal{D}}^2_t( \myvec{x}_i,\myvec{x}_j)=||{\mymat{\Psi}_t{(\myvec{x}_i)}}-{\mymat{\Psi}_t{(\myvec{x}_j)}}||^2=||\myvec{P}_{i,:}-\myvec{P}_{j,:}||^2_{\tiny\mymat{W}^{-1}}},}\] where \(\mymat{W}\) is a diagonal matrix with elements \(W_{i,i}=\frac{D_{i,i}}{\sum_{i=1}^M D_{i,i}}\). Thus, the DM embedding is distance preserving, meaning that neighboring points in the high-dimensional space are embedded close to each other in the diffusion coordinates.
# Data Fusion {#SecMulti}
Many physical phenomena are sampled using multiple types of sensing devices. Each sensor provides a noisy measurement of a latent parameter of interest. Data fusion is the process of incorporating multiple observation of the same data points to find a more coherent and accurate representation.\
**Problem Formulation:** Given multiple sets of data points \(\mymat{X}^l\text{ }, l=1...,L\). Each view is a high dimensional dataset \({\mymat{X}^l = \lbrace{ { \myvec{x}_1^l,\myvec{x}_2^l, \myvec{x}_3^l...,\myvec{x}_M^l }} \rbrace, \myvec{x}_i^l\in {\mathbb{R}^{D}} }\). Find a reliable low dimensional representation \(\myvec{\Psi}(\myvec{X}^1...,\myvec{X}^L)\in \mathbb{R}^d\).
## Multi-View Diffusion Maps (Multi-View DM) {#Sec2}
An approach for fusion kernel matrices in the spirit of DM framework was presented in. The idea is to enforce a random walk model based on the kernels that model each view by restraining the random walker to "hop" between views in each time step. The construction requires to compute a Gaussian kernel for each view \[\label{EQK}{ K^l_{i,j}=exp\{ {-\frac{||\myvec{x}^l_i-\myvec{x}^l_j||^2}{2 \sigma_l^2}\}},\text{ } l=1...,L },\] then the multi-view kernel is formed by the following matrix \[\label{EQKMAT} \mymat{\widehat{K}}= \begin{bmatrix} \mymat{0}_{M \times M} & {\mymat{K}^1\mymat{K}^2}& {\mymat{K}^1\mymat{K}^3}&...& {\mymat{K}^1\mymat{K}^p} \\ \mymat{K}^2\mymat{K}^1 & \mymat{0}_{M \times M} & {\mymat{K}^2\mymat{K}^3}&...& {\mymat{K}^2\mymat{K}^p}\\ \mymat{K}^3\mymat{K}^1 & {\mymat{K}^3\mymat{K}^2} & \mymat{0}_{M \times M} &...& {\mymat{K}^3\mymat{K}^p}\\:&:&:&...&:\\\mymat{K}^p\mymat{K}^1 & {\mymat{K}^p\mymat{K}^2} & {\mymat{K}^p\mymat{K}^3} &...& {\mymat{0}_{M \times M}}. \end{bmatrix}.\] Next, re-normalizing using the diagonal matrix \(\mymat{\widehat{D}}\) where \( {\widehat{D}}_{i,i}=\underset{j}{\sum}{{\widehat{K}}_{i,j}} \), the normalized row-stochastic matrix is defined as \[\label{phat} \mymat{\widehat{P}}={\mymat{\widehat{D}}}^{-1}\mymat{\widehat{K}}, ~~~ {\widehat{P}}_{i,j}=\frac{{{\widehat{K}}_{i,j}} }{\widehat{D}_{i,i} },\] where the \(m,l\) block is a square \(M\times M\) matrix located at\
\([1+(m-1)M,1+(l-1)M], l=1...,L\). This block describes the probability of transition between view \(\mymat{X}^m\) and \(\mymat{X}^l\). The multi-view DM representation for \(\mymat{X}^l\) is computed by \[\label{Map1} { \myvec{\widehat{\Psi}}_t{(\myvec{x}^l_i)}: \myvec{x}^l_i \longmapsto \begin{bmatrix} { \lambda_1^{t}\psi_1(i+\bar{l})}, {.} {.} {.}, {\lambda_{d}^{t}\psi_{d}(i+\bar{l})} \end{bmatrix}^T \in{\mathbb{R}^{d}} },\] where \(\bar{l}=(l-1)\cdot M\). The final low dimensional representation is defined by a concatenation of all low dimensional multi-view mappings \[\label{eq:MVDMrep} \vec{\myvec{\Psi}}(\vec{\myvec{X}})= {[ \myvec{\widehat{\Psi}}{(\myvec{X}^1)}},\myvec{\widehat{\Psi}}{(\myvec{X}^2)}...,\myvec{\widehat{\Psi}}{(\myvec{X}^L)}].\]
## Alternative Methods {#SECALT}
Here we provide a brief description of several methods for fusing the views before the application of a spectral decomposition.\
**Kernel Product (KP)**: Multiplying the kernel matrices element wise \(\mymat{K}^{{\circ}}\defeq \mymat{K}^1 \circ \mymat{K}^2 \circ...\circ \mymat{K}^L\), \({K}_{ij}^{\circ}\defeq {K}_{ij}^1 \cdot {K}_{ij}^2\cdot...\cdot {K}_{ij}^L\), then normalizing by the sum of rows. The resulting row stochastic matrix is denoted as \(\mymat{P}^{{\circ}}\). This kernel corresponds to the approach in.\
**Kernel Sum (KS)**: Defining the sum kernel \(\mymat{K}^{+}\defeq\sum^L_{l=1}\mymat{K}^l\). Normalizing the sum kernel by the sum of rows, to compute \(\mymat{P}^{+}\). This random walk sums the step probabilities from each view. This approach is proposed in.\
**Kernel Canonical Correlation Analysis (KCCA)**: This method detailed in extend the well know Canonical Correlation Analysis (CCA).Two kernels \(\myvec{K}^1 \text{and }\myvec{K}^2\) are constructed in each view as in Eq. ([\[EQK\]](#EQK){reference-type="ref" reference="EQK"}) and the canonical vectors \(\myvec{v}_1 \text{and }\myvec{v}_2\) are computed by solving the following generalized eigenvalue problem \[\label{eq:KCCA} \begin{bmatrix} \mymat{0}_{M \times M} & {\mymat{K}^1\cdot \mymat{K}^2} \\ {{\mymat{K}^2\cdot \mymat{K}^1}} & \mymat{0}_{M \times M} \end{bmatrix} \begin{pmatrix} {\myvec{v}_1} \\ {\myvec{v}_2} \end{pmatrix}= \rho \cdot \begin{bmatrix} (\mymat{K}^1+\gamma\myvec{I})^2 & {\mymat{0}_{M \times M}} \\ {\mymat{0}_{M \times M}} &(\mymat{K}^2+\gamma\myvec{I})^2 \end{bmatrix} \begin{pmatrix} {\myvec{v}_1} \\ {\myvec{v}_2} \end{pmatrix},\] where \(\gamma \myvec{I}\) are regularization terms which guarantee that the matrices \((\mymat{K}^1+\gamma\myvec{I})^2\) and \((\mymat{K}^2+\gamma\myvec{I})^2\) are invertible.
# Seismic Data Set {#sec:Data}
The data set that is used for demonstrating the proposed kernel based approaches includes \(2023\) explosions and \(105\) earthquakes. \(1654\) of the explosions occurred at the Shidiya phosphate quarry in the Southern Jordan between the years \(2005\)-\(2015\) (see a map of the region in Figure [\[fig:QuarryLocationsA\]](#fig:QuarryLocationsA){reference-type="ref" reference="fig:QuarryLocationsA"}). These events were reported by the Israel National Data Center at the Soreq Nuclear Research Center with magnitudes \(2\leq \text{ML}\leq3\) seismic. The rest of the events were taken from the seismic catalog of the Geophysical Institute of Israel between the years \(2004\)-\(2014\). All events were reported in Israel between latitudes \(29^{\circ}\)N-\(32.5^{\circ}\)N and longitudes \(34.2^{\circ}\)E-\(35.7^{\circ}\)E with duration magnitudes Md \(\ge 2.5\). Most of the earthquakes in the dataset occurred in the Dead Sea transform. The dataset includes the February 11, 2004 earthquake with the duration magnitude of Md = 5.1. This was the strongest event in this area since 1927. Twelve aftershocks that are included in the dataset are associated with this main shock. The majority of the explosions in the dataset are ripple-fire query blasts. Moreover, the dataset consists of several one shot explosions, for instance, two experimental underwater explosions in the Dead Sea and surface and near-surface experimental explosions at the Oron quarry and at the Sayarim Military Range in the Negev desert. The dataset consists of seismogram recordings from the HRFI (Harif) station in Israel. The station is part of the Israel National Seismic Network. It is equipped with a three component broad band STS-2 seismometer and a Quanterra data logger. The seismograms are sampled at a frequency of \(40\) Hz. Waveform segments of \(2.5\)-minutes (6000 samples) have been selected for every event. In each waveform, the first P phase onsets reside \(30\) seconds after the beginning of each waveform. Figure [\[fig:QuarryLocationsA\]](#fig:QuarryLocationsA){reference-type="ref" reference="fig:QuarryLocationsA"} displays the events on the regional map.
# Seismic Preprocessing and Feature Extraction Methods {#sec:Pre}
This section provides background on typical methods that are used for seismic signal processing as well as the description of the feature extraction method that was applied here. First, the STA/LTA detector is reviewed. Next, we describe how the alignment between the different waveforms was implemented. Last, the feature extraction step, which results in a time-frequency representation of the seismic signal, is described.
## Short and Long Time Average (STA/LTA) {#STA}
Detection of seismic signal embedded in the background noise is a classical problem in the signal processing theory. In the context of statistical decision theory it may be formulated as a choice between two alternatives: a waveform contains solely the noise or it contains a signal of interest superimposed on the noise. The STA/LTA trigger is a most widely accepted detection algorithm in seismology. It relies on the assumption that a signal is characterize by a concentration of higher energy level compared with the energy level of the noise. This is done by comparing short-time energy average to a long-time energy average using a Short Time Average/ Long Time Average (STA/LTA) detector. Usually a band-pass filter is applied before the STA/LTA test. Given a time signal \(\myvec{y}(n)\) the ratio \(R(i)\) is computed at each time instance \(i\) is computed as follows \[\label{eq:STALTA} R(i)=\frac{L\cdot[\overset{i+S}{\underset{{j=i}}{\sum}}y^2(j)]}{S\cdot(\overset{i+L}{\underset{{j=i}}{\sum}}y^2(j))},\] where \(L \gg S\) are the number of samples used for the long and short average correspondingly. The ratio \(R(i)\) is compared to a threshold \(\delta\) to identify time windows suspected as seismic events.
## Seismic Event Alignment
All waveform segments in the dataset were extracted according to the first P phase onset time. Those onset times were manually picked by the analysts. However, our selective waveform inspection showed that the P onsets often have actual offsets of several seconds, sometimes even of ten seconds. In order to increase the accuracy of the alignment, Algorithm [\[alg:Trigger\]](#alg:Trigger){reference-type="ref" reference="alg:Trigger"} is proposed to detect the first P onsets.
[\[alg:Trigger\]]{#alg:Trigger label="alg:Trigger"}
The bins are normalized such that the sum of energy in every frequency band is equal to \(1\). The resulted sonogram is denoted by \(\myvec{S}(k,t)\), where \(k\) is the frequency band number, and \(t\) is the time window number. Finally, we transpose the sonogram matrix into a Sonovector \(\myvec{x}\) by concatenating the columns such that \[\label{eq:Sonovec} \myvec{x}=\myvec{S}(:).\]
[\[alg:SeismicDM\]]{#alg:SeismicDM label="alg:SeismicDM"}
For this study 602 seismograms of explosions are used. The explosions occurred in 5 quarry clusters (see Table [2](#Table:Loc){reference-type="ref" reference="Table:Loc"} and Figure [\[fig:QuarryLocations\]](#fig:QuarryLocations){reference-type="ref" reference="fig:QuarryLocations"}) and the label data was taken from seismic catalogs. It should be noted that the quarry clusters may include several neighboring quarries and the quarry area may be of several kilometers (like Rotem) or more than ten kilometers (like Shidiya). Moreover, the precise ("ground truth") location for most of explosions inside a quarry are not known. We estimate that the hypo-center accuracy in the used seismic catalogs is about a few kilometers for the explosions in Israel and it is more than ten kilometers for the explosions in Jordan, which are located outside the Israeli seismic network. The mean latitude and longitude are computed for the explosions belonging to each cluster and referred them to the nearby quarry (see Table [2](#Table:Loc){reference-type="ref" reference="Table:Loc"}). The application of Algorithm [\[alg:SeismicDM\]](#alg:SeismicDM){reference-type="ref" reference="alg:SeismicDM"} yields a low dimensional representation of the seismic recordings. An example of a 3-dimensional single view DM mapping is presented in Figure [\[fig:QuarriesMapZ\]](#fig:QuarriesMapZ){reference-type="ref" reference="fig:QuarriesMapZ"}. The mapping is followed by a classification step that is performed based on a 1-fold cross validation using K-NN with \(K=3\). The accuracy of the classification is presented in Figure [\[fig:ClassifyQuarry\]](#fig:ClassifyQuarry){reference-type="ref" reference="fig:ClassifyQuarry"}. The multi-view approach shows a peak performance of 85% of correct classification rate.
## Location Estimation
The following case study demonstrates how the diffusion coordinates extract underlying physical properties of the sampled signal. In particular we show that the low dimensional representation that is generated by diffusion maps organizes the events with respect to their source location, even though this was not an input parameter of the algorithm. The original high-dimensional space holds the sonogram of each event. Nearly co-located events with the similar source mechanisms and magnitudes should have a similar time-frequency content and, consequently, have similar sonograms. Therefore, we expect them to lie close to each other in the high dimensional space. The diffusion distance, which is the metric that is preserved in DM, embeds the data while keeping its geometrical structure. Thus, physical properties (such as the source location) that characterize the sonogram and therefore define the geometric structure of the points in the high-dimensional space, are preserved in low-dimensional DM embedding. Note that such a geometry preserving metric does not exist in linear dimensionality reduction methods like PCA. The dataset for this study includes 352 explosions that occurred in 4 quarry clustering Israel out 5 clusters above. The explosions in Jordan were removed since they are located at a large distance from the HRFI station. We show that the location of seismic events can be evaluated from the DM embedding coordinates. A similar evaluation based on a linear projection that was calculated with PCA yields a less accurate correlation to the events' true location. Figure [\[fig:MapLoc\]](#fig:MapLoc){reference-type="ref" reference="fig:MapLoc"} (top image) displays the longitude and latitude coordinates of catalog locations of the events. These are the source locations of the seismic events. The points are colored by distance in kilometers from HRFI station. The middle and bottom images of Figure [\[fig:MapLoc\]](#fig:MapLoc){reference-type="ref" reference="fig:MapLoc"} present the two-dimensional PCA and DM embeddings of the dataset, respectively. It is clearly evident that the DM (bottom image in Figure [\[fig:MapLoc\]](#fig:MapLoc){reference-type="ref" reference="fig:MapLoc"}) representation has captured the location variability, while in the PCA representation this intrinsic factor is less obvious (middle image in Figure [\[fig:MapLoc\]](#fig:MapLoc){reference-type="ref" reference="fig:MapLoc"}). In the DM embedding, the clusters are well separated with respect to the event's location. In PCA the separation is not as clear, meaning that the low-dimensional PCA representation does not reveal this property. The Pearson correlation coefficients between first two diffusion coordinates and relative latitude and longitude are 0.82 and 0.77 for both dimensions respectively. The Pearson correlation coefficients between first two principle and relative latitude and longitude are 0.56 and 0.39 respectively.
## Detecting Anomalous Events
This case study demonstrates the diffusion representation's ability to detect anomalous events among set of events at specific site. When two events are nearly co-located, have close magnitudes but with different source mechanisms, then their sonograms should be quite different as well. Ripple-fire explosions are part of routine mining production cycles at the Oron phosphate quarry in Israel. In July 2006, three experimental one shot explosions were conducted by the Geophysical Institute of Israel at the Oron quarry. Our goal is to distinguish between the one shot explosions and the ripple-fire quarry blasts. This is not a trivial task, as all the events were conducted at very close distances.
[\[alg:KNNanomaly\]]{#alg:KNNanomaly label="alg:KNNanomaly"}
:::
To remove the variability created by the location of the events, 98 blast from a small region surrounding the ground truth location of the experimental explosions as reported in are used. Algorithm [\[alg:SeismicDM\]](#alg:SeismicDM){reference-type="ref" reference="alg:SeismicDM"} is applied and a mapping extracted from the Z-channel is used to find the suspected anomalies. The diffusion maps embedding is presented in Figure [\[fig:Anomalies\]](#fig:Anomalies){reference-type="ref" reference="fig:Anomalies"}. The three anomaly points are colored in blue, they are clearly separated from the main cluster. The anomalies are automatically identified using Algorithm [\[alg:KNNanomaly\]](#alg:KNNanomaly){reference-type="ref" reference="alg:KNNanomaly"} with \(K=4\) and a threshold set as four times the median of all distances \(\hat{D}_i,i=1...,M\). The average K-NN distance for the 98 blasts is presented in Figure [\[fig:AverageKNN\]](#fig:AverageKNN){reference-type="ref" reference="fig:AverageKNN"}. The four events that were suspected as anomalies include the three experimental explosions (which are described in ).
# Conclusion {#sec:Future}
In this paper, we have adapted a multi-view manifold learning framework for fusion of seismic recordings and for low-dimensional modeling. The abilities of kernel fusion methods for extracting meaningful seismic parameters were demonstrated on various case studies. Various algorithms for classification of seismic events type, location estimation and anomaly detection were presented. These algorithms can be used as decision support tools for analysts who need to determine the source, location and type of recorded seismic events. Correct classification of events results in improved and more accurate seismic bulletins. The proposed method is model free, thus it does not require knowledge of physical parameters. The underlying physical parameters are revealed by the diffusion maps and multi-view constructions. This type of kernel based sensor fusion is new in seismic signal processing and it overcomes some of the limitation of traditional model based fusion methods. | {'timestamp': '2017-06-07T02:06:35', 'yymm': '1706', 'arxiv_id': '1706.01750', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01750'} |
null | null |
null | null |
# Introduction {#intro}
The central goal of the beam energy scan (BES) program at RHIC is to seek evidence for the existence of a critical point in the phase diagram of strong-interaction matter. The hope is to detect this postulated second order phase transition point through the analysis of higher order cumulants of net charge fluctuations. Maxima of cumulants of net charge fluctuations, e.g. the \(2^{nd}\) and \(4^{th}\) order cumulants, trace the chiral crossover transition line at small values of the baryon chemical potential and diverge at a critical point. In heavy ion experiments the observed net charge fluctuations are expected to reflect thermal conditions at the time of chemical-freeze out of various hadron species. If this freeze-out happens close to the pseudo-critical line for the chiral transition of QCD, where thermal fluctuations are large, the measured fluctuations have a chance to be indicative for the divergent fluctuations that will appear at a critical point in the QCD phase diagram. A crucial anchor point for this scenario is to establish the relation between freeze-out and the QCD chiral transition at small or even vanishing net baryon chemical potential. In this case reliable theoretical calculations, based on Taylor series expansions in lattice QCD, exist and can be confronted with experimental findings at the LHC as well as the highest beam energies at RHIC. In experiments at the LHC one can analyze moments of charge fluctuations at almost vanishing baryon chemical potential (\(\mu_B\)) which allows a direct comparison with lattice QCD calculations performed at \(\mu_B=0\).
# Thermal conditions at vanishing net baryon number
The basic bulk thermodynamic observables, pressure (\(P\)), energy (\(\epsilon\)) and entropy (\(s\)) density of strong-interaction matter at vanishing baryon chemical potential, have been calculated in lattice QCD (for a recent review see: ). These calculations have recently been extended to non-vanishing baryon number densities using analytic continuation of calculations performed at imaginary values of \(\mu_B\) and Taylor expansions in \(\mu_B\). In Fig. [\[fig:phasediagram\]](#fig:phasediagram){reference-type="ref" reference="fig:phasediagram"} (left) we show results for lines of constant \(P\), \(\epsilon\) and \(s\), in the \(T\)-\(\mu_B\) plane (phase diagram) obtained from a Taylor series up to \({\cal O}(\mu_B^4)\). Lines are drawn for three values of these observables in the crossover region for the QCD chiral transition, which is well characterized by the current uncertainty on the chiral transition temperature, \(T_c=154(9)\) MeV, at \(\mu_B=0\). As can be seen in Fig. [\[fig:phasediagram\]](#fig:phasediagram){reference-type="ref" reference="fig:phasediagram"} (right) in this temperature interval the energy density changes by about a factor three, \(\epsilon_c = (0.34\pm 0.16)\) GeV/fm\(^3\). Also shown in Fig. [\[fig:phasediagram\]](#fig:phasediagram){reference-type="ref" reference="fig:phasediagram"} (left) are experimental results for freeze-out parameters determined by the ALICE Collaboration at the LHC and the STAR Collaboration from the BES at RHIC by comparing measured particle yields with predictions from a statistical hadronization model, which utilizes the thermodynamics of a hadron resonance gas. Obviously, there is a significant difference in the determination of the freeze-out temperature at \(\mu_B\simeq 0\). While the ALICE result for the freeze-out temperature (\(T_f\)) agrees well with the central value of the pseudo-critical temperature (\(T_c\)), the STAR results favor a larger value, \(T_f\sim 165\) MeV, which is close to the hadronization temperature obtained by Becattini et al.. A 10 MeV accuracy for the determination of the freeze-out temperature, which anyhow is not considered to be a temperature uniquely defined for all particle species, but rather a statistical average, may be considered to be appropriate for many purposes. However, in the search for evidence for a critical point such a difference has substantial consequences for expected properties of net charge fluctuations as the size of the critical region, in which charge fluctuations may become large, may well be only of that order . A 10 MeV difference between \(T_c\) and \(T_f\) thus may decide whether or not freeze-out happens in the critical region. Cumulants of net charge fluctuations and correlations among fluctuations of different conserved charges, i.e. baryon number (\(B\)), electric charge (\(Q\)) and strangeness (\(S\)) can be obtained as derivatives of the logarithm of the QCD partition function, \[\chi_{n}^{X} = \left. \frac{1}{VT^3} \frac{\partial^{n} \ln Z(V,T,\vec{\mu})}{\partial X^n} \right|_{\vec{\mu}=0}\;\; ,\;\; \chi_{nm}^{XY} = \left. \frac{1}{VT^3} \frac{\partial^{(n+m)} \ln Z(V,T,\vec{\mu})}{\partial X^n\ \partial Y^m} \right|_{\vec{\mu}=0}\;\; ,\;\; X,\ Y=B,\ Q,\ S \;\;, \label{cumulants}\] with \(\vec{\mu} \equiv (\mu_B,\ \mu_Q,\ \mu_S)\) denoting the three chemical potentials connected with the conserved charges. As can be seen in Fig. [\[fig:chiratios\]](#fig:chiratios){reference-type="ref" reference="fig:chiratios"} ratios of cumulants of net charge fluctuations change rapidly in the crossover region. The ratio of \(4^{th}\) and \(2^{nd}\) order net baryon-number fluctuations, \(\chi_4^B/\chi_2^B\), changes from almost unity at \(T=145\) MeV to about 0.5 at \(T=165\) MeV (see Fig. [\[fig:chiratios\]](#fig:chiratios){reference-type="ref" reference="fig:chiratios"} (left)). In the temperature range identified by ALICE as the freeze-out region, \(T_f=156(2)\) MeV, the ratio is \(\chi_4^B/\chi_2^B\simeq 0.75\). The situation is similar for ratios of conserved charge correlations. For the ratio of \(4^{th}\) and \(2^{nd}\) order cumulants characterizing correlations between net strangeness and net baryon-number fluctuations one finds \(\chi_{31}^{BS}/\chi_{11}^{BS}\simeq 0.74\) at \(T=155\) MeV, while this ratio drops to \(\sim 0.36\) at \(T=165\) MeV. This is in contrast to HRG model calculations with point-like non-interacting hadrons, where these ratios are unity irrespective of the particle content in the hadron spectrum. Also on the level of second order cumulants, which currently get measured by the ALICE Collaboration, differences between HRG model calculations and QCD results are already significant. In Fig. [\[fig:chiratios\]](#fig:chiratios){reference-type="ref" reference="fig:chiratios"} (right) we show the ratio, \(\chi_{11}^{BQ}/\chi_2^B\), which in resonance gas models for non-interacting, point-like hadrons has the interpretation of the relative contribution of charged baryons to the total baryon contribution of the pressure. As can be seen at \(T=165\) MeV the ratio \(\chi_{11}^{BQ}/\chi_2^B\) is about 50% smaller than predicted by HRG model calculations based on the experimentally known hadron spectrum (PDG-HRG). Model calculations that include additional strange baryon resonances predicted in QCD motivated quark model calculations (QM-HRG) provide a better approximation to QCD calculations but still over-predict the ratio \(\chi_{11}^{BQ}/\chi_2^B\) for \(T\raise0.3ex\hbox{\)>\(\kern-0.75em\raise-1.1ex\hbox{\)\sim\(}} 150\) MeV. A similar behavior has been found for the correlations between net baryon-number and net strangeness fluctuations, \(\chi_{11}^{BS}\), normalized to either \(\chi_2^B\) or \(\chi_2^S\). This may be taken as evidence for the importance of additional strange baryon resonance contributing significantly to the thermodynamics close to the chiral transition. There is, of course, also the possibility that additional interactions among hadrons need to be taken into account in order to improve agreement between QCD and hadron gas calculations at low temperature. However, as is also evident from Fig. [\[fig:chiratios\]](#fig:chiratios){reference-type="ref" reference="fig:chiratios"}, it seems that neither taking into account repulsive interactions through the finite volume of baryons nor an additional attractive contribution through a van der Waals interaction (vdW-HRG) can improve the validity range of HRG model based calculations.
# Cumulants of conserved charge fluctuations at non-zero net baryon number
Ratios of \(4^{th}\) and \(2^{nd}\) order cumulants such as \(\chi_4^B/\chi_2^B\) (Fig. [\[fig:chiratios\]](#fig:chiratios){reference-type="ref" reference="fig:chiratios"} (left)) or \(\chi_{31}^{BX}/\chi_{11}^{BX}\), \(X=Q,\ S\), also give the leading \({\cal O}(\mu_B^2)\) corrections to the \(2^{nd}\) order cumulants themselves. For \(\mu_Q=\mu_S=0\) one has \[\chi_{2}^{B}(T,\mu_B) = \chi_{2}^{B} + \frac{1}{2}\chi_{4}^{B} \left( \frac{\mu_B}{T} \right)^2+ {\cal O}(\mu_B^4)\;\;\;\; ,\;\;\;\; \chi_{11}^{BX}(T,\mu_B) = \chi_{11}^{BX} + \frac{1}{2}\chi_{31}^{BX} \left( \frac{\mu_B}{T} \right)^2+ {\cal O}(\mu_B^4) \;\;,\;\; X=Q, \; S \;. \label{expansion}\] As these ratios are smaller than the corresponding HRG values, this also means that differences between lattice QCD results for these cumulants and HRG model calculations increase with \(\mu_B\). Some results for \(\chi_{11}^{BS}\) and \(\chi_{11}^{BQ}\) at three values of the temperature are shown in Fig. [\[fig:chi42\]](#fig:chi42){reference-type="ref" reference="fig:chi42"}. Obviously deviations from HRG model calculations become large for \(\mu_B/T \raise0.3ex\hbox{\)>\(\kern-0.75em\raise-1.1ex\hbox{\)\sim\(}} 1.5\), which in the BES at RHIC corresponds to \(\sqrt{s_{NN}} \raise0.3ex\hbox{\)<\(\kern-0.75em\raise-1.1ex\hbox{\)\sim\(}} 15\) GeV. The situation is similar for ratios of cumulants of net baryon-number fluctuations. For instance, \[\frac{\chi_4^B(T,\mu_B)}{\chi_2^B(T,\mu_B)} = \frac{\chi_4^B}{\chi_2^B}\left( 1 + \frac{1}{2} \left( \frac{\chi_6^B}{\chi_4^B}-\frac{\chi_4^B}{\chi_2^B} \right) \left( \frac{\mu_B}{T} \right)^2 + {\cal O}(\mu_B^4) \right) \;\;\; {\rm for} \;\;\; \mu_Q=\mu_S=0\;. \label{B42}\] is unity in HRG model calculations with non-interacting, point-like hadrons for all values of \(\mu_B\). In QCD, however, the \({\cal O}(\mu_B^2)\) expansion coefficient in Eq. [\[B42\]](#B42){reference-type="ref" reference="B42"} is negative for \(T\raise0.3ex\hbox{\)>\(\kern-0.75em\raise-1.1ex\hbox{\)\sim\(}} 150\) MeV. This indicates that the ratio \(\chi_4^B(T,\mu_B) / \chi_2^B(T,\mu_B)\) becomes smaller with increasing \(\mu_B\). Even if the ratio is close to HRG model results at \(\mu_B=0\), it thus will further deviate from this with increasing \(\mu_B\). This is consistent with the trend found by STAR for the corresponding kurtosis ratio of net proton-number fluctuations, \(\kappa_P\sigma^2_P=\chi_4^P/\chi_2^P\).
# Conclusions
Many ratios of \(4^{th}\) and \(2^{nd}\) order cumulants of conserved charge fluctuations calculated in QCD agree with HRG model calculations within 25% or better at temperatures below \(T=155\) MeV. For observables including strangeness fluctuations this often requires to take into account contributions from additional strange hadrons not listed in the Particle Data Tables. At \(T=165\) MeV HRG model calculations often deviate from QCD results by more than 50% even for observables that only involve quadratic charge fluctuations. This work was supported through Contracts No. DE-SC001270 with the U.S. Department of Energy and No. 05P15PBCAA with the German Bundesministerium für Bildung und Forschung. | {'timestamp': '2017-06-07T02:03:37', 'yymm': '1706', 'arxiv_id': '1706.01620', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01620'} |
# Introduction
Establishing causal relations between variables from observation of their behaviour in time is central to scientific investigation and it is at the core of data-science where these causal relations are the basis for the construction of useful models and tools capable of prediction. The capability to predict (future) outcomes from the analytics of (past) input data is crucial in modeling and it should be the main property to take into consideration in model selection, when the validity and meaningfulness of a model is assessed. From an high-level perspective, we can say that the whole scientific method is constructed around a circular procedure consisting in observation, modelling, prediction and testing. In such a procedure, the accuracy of prediction is used as a selection tool between models. In addition, the principle of parsimony favours the simplest model when two models have similar predictive power. The scientific method is the rational process that, for the last 400 years, has mostly contributed to scientific discoveries, technological progresses and the advancement of human knowledge. Machine learning and data-science are nowadays pursuing the ambition to mechanize this discovery process by feeding machines with data and using different methodologies to build systems able to make models and predictions by themselves. However, the automatisation of this process requires to identify, without the help of human intuition, the relevant variables and the relations between these variables out of a large quantity of data. Predictive models are methodologies, systems or equations which identify and make use of such relations between sets of variables in a way that the knowledge about a set of variables provides information about the values of the other set of variables. This problem is intrinsically high-dimensional with many input and output data. Any model that aims to explain the underlying system will involve a number of elements which must be of the order of magnitude of the number of relevant relations between the system's variables. In complex systems, such as financial markets or the brain, prediction is probabilistic in nature and modeling concerns inferring the probability of the values of a set of variables given the values of another set. This requires the estimation of the joint probability of all variables in the system and, in complex systems, the number of variables with potential macroscopic effects on the whole system is very large. This poses a great challenge for the model construction/selection and its parameter estimation because the number of relations between variables scales with-at least-the square of the number of variables but, observation window, the amount of information gathered from such variables scales-at most-linearly with the number of variables. For instance, a linear model for a system with \(p\) variables requires the estimation from observation of \(p(p+1)/2\) parameters (the distinct elements of the covariance matrix). In order to estimate \(\mathcal O(p^2)\) parameters one needs a comparable number of observations requiring time series of length \(q \sim p\) or larger to gather a sufficient information content from a number of observations which scales as \(p \times q \sim \mathcal O(p^2)\). However, the number of parameters in the model can be reduced by considering only \(\mathcal O(p)\) out of the \(\mathcal O(p^2)\) relations between the variables reducing in this way the required time series length to \(\mathcal O(p)\). Such models with reduced numbers of parameters are referred in the literature as sparse models. In this paper we consider two instances of linear sparse modelling: Glasso which penalizes non-zero parameters by introducing a \(\ell_1\) norm penalization and LoGo which reduces the inference network to an \(\mathcal O(p)\) number of links selected by using information filtering networks . The results from these two sparse models are compared with the \(\ell_2\) norm penalization (non-sparse) ridge model. This paper is an exploratory attempt to map the parameter-regions of time series length, number of variables, penalization parameters and kinds of models to define the boundaries where probabilistic models can be reasonably constructed from the analytics of observation data. In particular, we investigate empirically, the true link retrieval performances in the region of short time-series and large number of variables which is the most critical region--and the most interesting--in many practical cases. Results are reported for artificially generated time series from an autoregressive model of \(p=100\) variables and time series lengths \(q\) between 10 and 20,000 data points. Robustness of the results has been verified over a wider range of \(p\) from 20 to 200 variables. Our results demonstrate that sparse models are superior in retrieving the true causality structure for short time series. Interestingly, this is despite considerable inaccuracies in the inference network of these sparse models. We indeed observe that statistical validation of causality is crucial in identifying the true causal links, and this identification is highly enhanced in sparse models. The paper is structured as follows. In section [\[s.definitions\]](#s.definitions){reference-type="ref" reference="s.definitions"} we briefly review the basic concepts of mutual information and conditional transfer entropy and their estimation from data that will then be used in the rest of the paper. We also introduce the concepts of sparse inverse covariance, inference network and causality networks. Section [3](#s.causalNet){reference-type="ref" reference="s.causalNet"} concerns the retrieval of causality network from the computation and statistical validation of conditional transfer entropy. Results are reported in Section [4](#s.results){reference-type="ref" reference="s.results"} where the retrieval of the true causality network from the analytics of time series from an autoregressive process of \(p=100\) variables is discussed. Conclusions and perspectives are given in Section [5](#s.conclusions){reference-type="ref" reference="s.conclusions"}.
# Estimation of conditional transfer entropy from data {#s.conditionslSigma}
[\[s.definitions\]]{#s.definitions label="s.definitions"} In this paper causality is quantified by means of statistically-validated transfer entropy. Transfer entropy \(T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j})\) quantifies the amount of uncertainty on a random variable, \(\mathbf{ Z_j}\), explained by the past of another variable, \(\mathbf{Z_i}\) conditioned to the knowledge about the past of \(\mathbf{ Z_j}\) itself. Conditional transfer entropy, \(T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j} | \mathbf W)\), includes an extra condition also to a set variables \(\mathbf W\). These quantities are introduced in details in Appendix [\[cTE\]](#cTE){reference-type="ref" reference="cTE"} (see also ). Let us here just report the main expression for the conditional transfer entropy that we shall use in this paper: \[\begin{aligned}
\label{General_cTE} T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j}|\mathbf{ W}) & {}= H(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf j,t}^{lag},\mathbf{ W}_{t}\})-H(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf i,t}^{lag},\mathbf Z_{\mathbf j,t}^{lag},\mathbf{ W}_{t}\}) \;\;. \end{aligned}\] Where \(H(.|.)\) is the conditional entropy, \(\mathbf Z_{\mathbf j,t}\) is a random variable at time \(t\), whereas \(\mathbf Z_{\mathbf i,t}^{lag} =\{ \mathbf Z_{\mathbf i,t-1}...,\mathbf Z_{\mathbf i,t-\tau}\}\) is the lagged set of random variable '\(\mathbf i\)' considering previous times \(t-1...t-\tau\) and \(\mathbf{ W}_{t}\) are all other variables and their lags (see Appendix [\[cTE\]](#cTE){reference-type="ref" reference="cTE"}, Eq.[\[General_cTE1\]](#General_cTE1){reference-type="ref" reference="General_cTE1"}). In this paper we use Shannon entropy and restrict to linear modeling with multivariate normal setting (see Appendix [\[Shannon\]](#Shannon){reference-type="ref" reference="Shannon"}). In this context the conditional transfer entropy can be expressed in terms of the determinants of conditional covariances \(\mbox{det}( \mathbf \Sigma(.|.))\) (see Eq.[\[H2\]](#H2){reference-type="ref" reference="H2"} in Appendix [\[Shannon\]](#Shannon){reference-type="ref" reference="Shannon"}): \[\begin{aligned}
\label{General_cTE_Sigma} T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j}|\mathbf{ W}) & {}= \frac12 \log \mbox{det}\!\left( \mathbf \Sigma(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf j,t}^{lag},\mathbf{ W}_{t}\})\right)-\frac12 \log \mbox{det}\!\left( \mathbf \Sigma(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf j,t}^{lag},\mathbf Z_{\mathbf i,t}^{lag},\mathbf{ W}_{t}\})\right) \;\;. \end{aligned}\] Conditional covariances can be conveniently computed in terms of the inverse covariance of the whole set of variables \(\mathbf Z_t = \{ \mathbf Z_{k,t},\mathbf Z_{k,t-1}...\mathbf Z_{k,t-\tau}\}_{k=1}^p \in \mathbb R^{p\times(\tau+1)}\) (see Appendix [\[InverseCovJandSigma\]](#InverseCovJandSigma){reference-type="ref" reference="InverseCovJandSigma"}). Such inverse covariance matrix, \(\mathbf J\), represents the structure of conditional dependencies among all couples of variables in the system and their lags. Each sub-part of \(\mathbf J\) is associated with the conditional covariances of the variables in that part with respect to all others. In terms of \(\mathbf J\), the expression for the conditional transfer entropy becomes: \[\begin{aligned}
\label{TE20} T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j}|\mathbf{ W}) =-\frac12 \log\mbox{det}\! \left(\mathbf {J_{1,1}}-\mathbf {J_{1,2}} (\mathbf {J_{2,2}})^{-1} \mathbf {J_{2,1}})\right)+ \frac12 \log \mbox{det}\! \left (\mathbf{ J_{1,1} }\right ) \;\;. \end{aligned}\] where the indices '\(\mathbf 1\)' and '\(\mathbf 2\)' refer to sub-matrices of \(\mathbf J\) respectively associated with the variables \(\mathbf Z_{\mathbf j,t}\) and \(\mathbf Z_{\mathbf i,t}^{lag}\). Conditioning eliminates the effect of the other variables retaining only the exclusive contribution from the two variables in consideration. This should provide estimations of transfer entropy that are less affected by spurious effects from other variables. On the other hand, conditioning in itself can introduce spurious effects, indeed two independent variables can become dependent due to conditioning . In this paper we explore two extreme conditioning cases: i) condition to all other variables and their lags; ii) unconditioned. In principle, one would like to identify the maximal value of \(T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j} | \mathbf W)\) over all lags and all possible conditionings \(\mathbf W\). However, the use of multiple lags and conditionings increases the dimensionality of the problem making estimation of transfer entropy very hard especially when only a limited amount of measurements is available (i.e. short time-series). This is because the calculation of the conditional covariance requires the estimation of the inverse covariance of the whole set of variables and such an estimation is strongly affected by noise and uncertainty. Therefore, a standard approach is to reduce the number of variables and lags to keep dimensionality low and estimate conditional covariances with appropriate penalizers. An alternative approach is to invert the covariance matrix only locally on low dimensional sub-sets of variables selected by using information filtering networks and then reconstruct the global inversion by means of the LoGo approach. Let us here briefly account for these two approaches.
## Penalized inversions
The estimate of the inverse covariance is a challenging task to which a large body of literature has been dedicated. From an intuitive perspective, one can say that the problem lies in the fact that uncertainty is associated with nearly zero eigenvalues of the covariance matrix. Variations in these small eigenvalues have relatively small effects on the entries of the covariance matrix itself but have major effects on the estimation of its inverse. Indeed small fluctuations of small values can yield to unbounded contributions to the inverse. A way to cure such near-singular matrices is by adding finite positive terms to the diagonal which move the eigenvalues away from zero: \(\hat {\mathbf J} = \left((1-\gamma) \mathbf S +\gamma \mathbf I_N \right)^{-1}\), where \(\mathbf S = \mbox{Cov}(\mathbf Z)\) is the covariance matrix of the set of variables \(\mathbf Z \in \mathbb R^N\) estimated from data and \(\mathbf I_{N}\in \mathbb R^{N\times N}\) is the identity matrix (where \(N=p\times(\tau+1)\), see later). This is what is performed in the so-called ridge regression, also known as shrinkage mean-square-error estimator or Tikhonov regularization. The effect of the additional positive diagonal elements is quivalent to compute the inverse covariance which maximizes the log-likelihood: \(\log \mbox{det}( \hat {\mathbf J})-\mbox{tr}(\mathbf S\hat{\mathbf J})-\gamma ||\hat {\mathbf J} ||_2\), where the last term penalizes large off-diagonal coefficients in the inverse covariance with a \(\ell_2\) norm penalization. The regularizer parameter \(\gamma\) tunes the strength of this penalization. This regularization is very simple and effective. However, with this method insignificant elements in the precision matrix are penalized toward small values but they are never set to zero. By using instead \(\ell_1\) norm penalization: \(\log \mbox{det}( \hat {\mathbf J})-\mbox{tr}(\mathbf S\hat{\mathbf J})-\gamma ||\hat{\mathbf J} ||_1\), insignificant elements are forced to zero leading to a sparse inverse covariance. This is the so-called lasso regularization. The advantage of a sparse inverse covariance consists in the provision of a network Indeed, let us recall that zero entries in the inverse covariance are associated with couples of non-conditionally dependent variables.
## Information filtering network approach: LoGo
An alternative approach to obtain sparse inverse covariance is by using information filtering networks generated by keeping the elements that contribute most to the covariance by means of a greedy process. of this approach is that inversion is performed at local level on a small subsets of variables and then the global inverse is reconstructed by joining the local parts through the information filtering network. Because of this Local-Global construction this method is named LoGo. It has been shown that LoGo method yields to statistically significant sparse precision matrices that outperform the ones with the same sparsity computed with lasso method.
# Causality network reterival {#s.causalNet}
## Simulated multivariate autoregressive linear process
In order to be able to test if causality measures can retrieve the true causality network in the underlying process, we generated artificial multivariate normal time series with known sparse causality structure by using the following autoregressive multivariate linear process: \[\begin{aligned}
\label{Process} \mathbf Z_t = \sum_{\lambda=1}^\tau \mathbf A_\lambda \mathbf Z_{t-\lambda} + \mathbf U_t \end{aligned}\] where \(\mathbf A_\lambda \in \mathbb R^{p \times p}\) are matrices with random entries drawn from a normal distribution. The matrices are made upper diagonal (diagonal included) by putting to zero all lower diagonal coefficients and made sparse by keeping only a \(\mathcal O(p)\) total number of entries different from zero in the upper and diagonal part. \(\mathbf U_t \in \mathbb R^p\) are random normally distributed uncorrelated variables. This process produces autocorrelated, cross-correlated and causally dependent time series. We chose it because it is among the simplest processes that can generate this kind of structured datasets. The dependency and causality structure is determined by the non-zero entries of the matrices \(\mathbf A_\lambda\). The upper-triangular structure of these matrices simplify the causality structure eliminating causality cycles. Their sparsity reduces dependency and causality interactions among variables. The process is made autoregressive and stationary by keeping the eigenvalues of \(\mathbf A_\lambda\) all smaller than one in absolute value. For the tests we used \(\tau=5\), \(p=100\) and sparsity is enforced to have a number of links approximately equal to \(p\). We reconstructed the network from time series of different lengths \(q\) between 5 to 20,000 points. To test statistical reliability the process was repeated 100 times with every time a different set of randomly generated matrices \(\mathbf A_\lambda\). We verify that the results are robust and consistent by varying sample sizes from \(p=20\) to \(200\), by changing sparsity with number of links from \(0.5 p\) to \(5 p\) and for \(\tau\) from 1 to 10.
## Causality and inference network retrieval
We tested the agreement between the causality structure of the underlying process and the one inferred from the analysis of \(p\) time-series of different lengths \(q\), \(\mathbf Z_t \in \mathbb R^p\) with \(t=1..q\), generated by using Eq.[\[Process\]](#Process){reference-type="ref" reference="Process"}. We have \(p\) different variables and \(\tau\) lags. The dimensionality of the problem is therefore \(N = p \times (\tau+1)\) variables at all lags including zero. To estimate the inference and causality networks we started by computing the inverse covariance, \(\mathbf J \in \mathbb R^{ N \times N}\), for all variables at all lags \(\mathbf Z \in \mathbb R^{ N \times q}\) by using the following three different estimation methods:
- \(\ell_1\) norm penalization (Glasso );
- \(\ell_2\) norm penalization (ridge );
- information filtering network (LoGo ).
We retrieved the inference network by looking at all couples of variables, with indices \({\mathbf i} \in[1..,p]\) and \({\mathbf j} \in[1..,p]\), which have non-zero entries in the inverse covariance matrix \(\mathbf J\) between the lagged set of \(\mathbf j\) and the non-lagged \(\mathbf i\). Clearly, for the ridge method the result is a complete graph but for the Glasso and LoGo the results are sparse networks with edges corresponding to non-zero conditional transfer entropies between variables \(\mathbf i\) and \(\mathbf j\). For the LoGo calculation we make use of the regularizer parameter as a local shrinkage factor to improve the local inversion of the covariance of the 4-cliques and triangular separators (see ). We then estimated transfer entropy between couples of variables, \({\mathbf i} \rightarrow {\mathbf j}\) conditioned to all other variables in the system. This is obtained by estimating of the inverse covariance matrix (indicated with an 'hat' symbol) by using Eq.[\[TE20b\]](#TE20b){reference-type="ref" reference="TE20b"} (see Appendix [\[s.CTE\]](#s.CTE){reference-type="ref" reference="s.CTE"}) with: \[\begin{aligned}
\mathbf{Z_1} &= \mathbf{ Z_j}_{,t}\\ \nonumber \mathbf{Z_2} &=\{\mathbf{ Z_i}_{,t-1}...\mathbf{ Z_i}_{,t-\tau}\}\\ \nonumber \mathbf{Z_3} &= \{\mathbf{ Z_j}_{,t-1}...\mathbf{ Z_j}_{,t-\tau},\mathbf{ W}\} \;. \end{aligned}\] With \(\mathbf{ W}\) a conditioning to all variables \(\mathbf{Z}\) except \(\mathbf{Z_1},\mathbf{Z_2}\) and \(\{\mathbf{ Z_j}_{,t-1}...\mathbf{ Z_j}_{,t-\tau}\}\). The result is a \(p \times p\) matrix of conditional transfer entropies \(T(\mathbf{ Z_i}_{,t} \rightarrow \mathbf{ Z_j}_{,t})\). Finally, to retrieve the causality network we retained the network of statistically validated conditional transfer entropies only. Statistical validation was performed as follows.
## Statistical validation of causality
Statistical validation has been performed from likelihood ratio statistical test. Indeed, entropy and likelihood are intimately related: entropy measures uncertainty and likelihood measures the reduction in uncertainty provided by the model. Specifically, the Shannon entropy associated with a set of random variables, \(\mathbf{ Z_i}\), with probability distribution \(p(\mathbf{ Z_i})\) is \(H(\mathbf{ Z_i}) =-\mathbb E [ \log p(\mathbf{ Z_i}) ]\) (Eq.[\[entropy\]](#entropy){reference-type="ref" reference="entropy"}) whereas the log-likelihood for the model \(\hat p(\mathbf{ Z_i})\) associated with a set of independent observations \(\mathbf{ \hat Z}_{i,t}\) with \(t=1..q\) is \(\log \mathcal L(\mathbf{\hat Z}_{\mathbf i}) = \sum_{t=1}^q \log \hat p(\mathbf{\hat Z}_{i,t})\) which can be written as \(\log \mathcal L(\mathbf{\hat Z}_{\mathbf i}) = q \mathbb E_{\hat p} [ \log \hat p(\mathbf{ Z_i}) ]\). Note that \(q\) is the total available number of observations which, in practice, is the length of the time-series minus the maximum number of lags. It is evident from these expressions that entropy and the log-likelihood are strictly related trough this link might be non-trivial. In the case of linear modeling this connection is quite evident because the entropy estimate is \(H = \frac 12 (-\log |\mathbf{\hat J} | + p \log (2\pi) + p)\) and the log-likelihood is \(\log \mathcal L = \frac q2 ( \log |\mathbf{\hat J} |-Tr( \mathbf{\hat \Sigma} \mathbf{\hat J} )-p \log (2\pi))\). For the three models we study in this paper we have \(Tr( \mathbf{\hat \Sigma} \mathbf{\hat J} ) =p\) and therefore the log-likelihood is equal to \(q\) times the opposite of the entropy estimate. Transfer entropy, or conditional transfer entropy, are differences between two entropies: the one of a set of variables conditioned to their own past minus the one conditioned also to the past of another variable. This, in turns, is the difference of the unitary log-likelihood of two models and therefore it is the logarithm of a likelihood ratio. As Wilks pointed out the null distribution of such model is asymptotically quite universal. Following the likelihood ratio formalism, we have \(\lambda=q T\) and the probability of observing a transfer entropy larger than \(T\), estimated under null hypothesis, is given by \(p_v \sim 1-\chi_c^2(r qT,d)\) with \(r\simeq 2\) and \(\chi_c^2\) the chi-square the cumulative distribution function with \(d\) degrees of freedom which are the difference between the number of parameters in the two models. In our case the two models have respectively \(\tau (p_j^2+1)\) and \(\tau(p_j^2+1) + \tau (p_j \, p_i)\) parameters.
## Statistical validation of the network
The procedures described in the previous two subsections produce the inference network and causality network. Such networks are then compared with the known underlying network of true causalities in the underlying process which is defined by the non-zero elements in the matrices \(A_\lambda\) (see Eq.[\[Process\]](#Process){reference-type="ref" reference="Process"}). The overlapping between the retrieved links in the inference or causality networks with the ones in the true network underlying the process is an indication of a discovery of a true causality relation. However some discoveries can be obtained just by chance or some methodologies might discover more links only because they produce denser networks. We therefore tested the hypothesis that the matching links in the retrieved networks are not obtained just by chances by computing the null-hypothesis probability to obtain the same or a larger number of matches randomly. Such probability is given by the conjugate cumulative hypergeometric distribution for a number equal or larger than \(\mathrm{TP}\) of 'true positive' matching causality links between an inferred network of \(n\) links and a process network of \(K\) true causality links, from a population of \(p^2-p\) possible links: \[\begin{aligned}
\label{HyperTest} P(X \ge \mathrm{TP} | n, K, p)= 1-\sum_{k=0}^{\mathrm{TP}-1} {\frac {{\binom {K}{k}}{\binom {p^2-p-K}{n-k}}}{\binom {p^2-p}{n}}} \;\;\;. \end{aligned}\] Small values of \(P\) indicate that the retrieved \(\mathrm{TP}\) links out of \(K\) are unlikely to be found by randomly picking \(n\) edges from \(p^2-p\) possibilities. Note that in the confusion matrix notation we have with \(\mathrm{TP}\) number of true positives, \(\mathrm{FP}\) number of false positives, \(\mathrm{FN}\) number of false negatives and \(\mathrm{TN}\) number of true negatives.
# Results {#s.results}
## Computation and validation of conditional transfer entropies
By using Eq.[\[Process\]](#Process){reference-type="ref" reference="Process"} we generated 100 multivariate autoregressive processes with known causality structures. We here report results for \(p=100\) but analogous outcomes were observed for dimensionalities between \(p=20\) and \(200\) variables. Conditional transfer entropies between all couples of variables, conditioned to all other variables in the system, were computed by estimating the inverse covariances by using tree methodologies, ridge, lasso and LoGo and applying Eq.[\[TE20\]](#TE20){reference-type="ref" reference="TE20"}. Conditional transfer entropies were statistically validated with respect to null hypothesis (no causality) at \(p_v=1\%\) p-value. Results for Bonferroni adjusted p-value at 1% (i.e. \(p_v = 0.01/(p^2-p) \sim 10^{-6}\) for \(p=100\)) are reported in Appendix [\[BonferroniValidated\]](#BonferroniValidated){reference-type="ref" reference="BonferroniValidated"}. We also tested other values of \(p_v\) from \(10^{-8}\) to 0.1 obtaining consistent results. We observe that small \(p_v\) reduce the number of validated causality links but increase the chance that these links match with the true network in the process. Conversely large values of \(p_v\) increase the numbers of mismatched links but also of the true links discoveries. Let us note that here we use \(p_v\) as a thresholding criteria and we are not claiming any evidence of statistical significance of the causality. We assess the goodness of this choice a-posteriori by comparing the resulting causality network with the known causality network of the process.
## Statistical significance of the recovered causality network.
Results for the contour frontiers of significant causality links for the three models are reported in Fig.[\[f.significance\]](#f.significance){reference-type="ref" reference="f.significance"} for a range of time series with lengths \(q\) between 10 and 20,000 and regularizer parameters \(\gamma\) between \(10^{-8}\) and \(0.5\). Statistical significance is computed by using Eq.[\[HyperTest\]](#HyperTest){reference-type="ref" reference="HyperTest"} and results are reported for both \(P < 0.05\) and \(P < 10^{-8}\) (continuous and dotted lines respectively). As one can see, the overall behaviours for the three methodologies are little affected by the threshold on \(P\). We observe that LoGo significance region extends well beyond the Glasso and ridge regions. The value of the regularizer parameter \(\gamma\) affects the results for the three models in a different way. Glasso has a region in the plane \(\gamma-p/q\) where it has best performances (in this case it appears to be around \(\gamma \simeq 0.1\) and \(p/q \simeq 2.5\)). Ridge appears instead to be little affected with mostly constant performances across the range of \(\gamma\). LoGo has best performances for small, even infinitesimal, values of \(\gamma\). Indeed, differently from Glasso in this case \(\gamma\) does not control sparsity but instead acts as local shrinkage parameter. Very small values can be useful in some particular cases to reduce the effect of noise but large values have only the effect to reduce information.
## Inference network
We have so far empirically demonstrated that a significant part of the true causality network can be retrieved from the statistically validated network of conditional transfer entropies. Results depend on the choice of the threshold value of the \(p_v\) at which null hypothesis is rejected. We observed that lower \(p_v\) are associated with network with fewer true positives but also fewer false positives and conversely larger \(p_v\) yield to causality networks with larger true positives but also larger false positives. Let us here report on the extreme case of the inference network which contains all causality channels with no validation. For the ridge model this network is the complete graph with all variables connected to each-other. Instead, for Glasso and LoGo the inference network is sparse. Results are summarized in Table.[2](#tab:fraction_Inference){reference-type="ref" reference="tab:fraction_Inference"}. In terms of true positive rate we first notice that they are all larger than the ones in Table.[1](#tab:fraction){reference-type="ref" reference="tab:fraction"}. Indeed, the network of statistically validated conditional transfer entropies is a sub-network of the inference network. On the other hand we notice that the false positive fraction is much larger than the ones in Table.[2](#tab:fraction_Inference){reference-type="ref" reference="tab:fraction_Inference"}. Ridge network has a fraction of 1 because, in this case, the inference network is the complete graph. Galsso also contains a very large number of false positives reaching even 55 times the number of links in the true network and getting to lower fractions only from long time-series with \(q >1000\). These numbers also indicate that Galsso networks are not sparse. LoGo has a sparser and more significant inference network with smaller fractions of false positives which stay below \(5n\), which is anyway a large number of misclassification. Nonetheless, we observe that, despite such large fractions of \(\mathrm{FP}\), the discovered true positives are statistically significant.
:::
[\[tab:fraction_Unconditional\]]{#tab:fraction_Unconditional label="tab:fraction_Unconditional"}
## Unconditioned transfer entropy network
We last tested whether conditioning to the past of all other variables gives better causality network retrievals than the unconditioned case. Here, transfer entropy, \(T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j})\), is computed by using Eq.[\[TE20\]](#TE20){reference-type="ref" reference="TE20"} with \(\mathbf W=\emptyset\), the empty set. For the ridge case this unconditional transfer entropy depends only from the time-series, \(\mathbf{ Z}_{i,t}\), \(\{\mathbf{ Z}_{i,t-1}...,\mathbf{ Z}_{i,t-\tau}\}\) and \(\{\mathbf{ Z}_{j,t-1}...,\mathbf{ Z}_{j,t-\tau}\}\) (with \(\tau = 5\) in this case). Glasso and LoGo cases are instead hybrid because a conditional dependency has been already introduced in the sparse structure of the inverse covariance \(\mathbf J\) (the inference network). Results are reported in Tab.[3](#tab:fraction_Unconditional){reference-type="ref" reference="tab:fraction_Unconditional"} where we observe that these networks retrieve a larger quantity of true positives than the ones constructed from conditional entropy. However, the fraction of false positive is also larger than the ones in Tab.[1](#tab:fraction){reference-type="ref" reference="tab:fraction"} although it is smaller than what observed in the inference network in Tab.[2](#tab:fraction_Inference){reference-type="ref" reference="tab:fraction_Inference"}. Overall, these results indicate that conditioning is effective in discarding false positives.
## Summary of all results in a single ROC plot
In summary, we have investigated the networks associated with conditional transfer entropy, unconditional transfer entropy and inference for three models under a range of different parameters. In the previous sub-sections we have provided some comparisons between the performances of the three models in different ranges of parameters. Let us here provide a summary of all results within a single ROC plot. Figure [\[f.ROC\]](#f.ROC){reference-type="ref" reference="f.ROC"} reports the ROC values, for each model and each parameter combination, x-axis are false positive rates (\(\mathrm{FP}/m\)) and y-axis true positive rates (\(\mathrm{TP}/n\)). Each point is an average over 100 processes. Points above the diagonal line are associated to relatively well performing models with the upper left corner representing the point where models correctly discover all true causality links without any false positive. The plot reports with large symbols the cases for \(\gamma = 0.1\) and validation at p-value \(p_v=0.01\), which can be compared with the data reported in the tables. Overall from Tables [1](#tab:fraction){reference-type="ref" reference="tab:fraction"}, [2](#tab:fraction_Inference){reference-type="ref" reference="tab:fraction_Inference"}, [3](#tab:fraction_Unconditional){reference-type="ref" reference="tab:fraction_Unconditional"} and Fig. [\[f.ROC\]](#f.ROC){reference-type="ref" reference="f.ROC"} we conclude that all models obtain better results for longer time series and that conditional transfer entropy over-performs the unconditional counterparts (see, Tables [1](#tab:fraction){reference-type="ref" reference="tab:fraction"} and [3](#tab:fraction_Unconditional){reference-type="ref" reference="tab:fraction_Unconditional"} and the two separated ROC figures for conditional and unconditional transfer entropies reported in Fig.[\[f.ROCCoUc\]](#f.ROCCoUc){reference-type="ref" reference="f.ROCCoUc"} in appendix [\[A.Co.UC\]](#A.Co.UC){reference-type="ref" reference="A.Co.UC"}). In the range of short time series, when \(q\le p\), which is of interest for this paper, LoGo is the best performing model with better performances achieved for small \(\gamma \lesssim 10^{-4}\) and validation with small p-values \(p_v \lesssim 10^{-4}\). LoGo is consistently the best performing model also for longer time-series up to lengths of \(q \sim1000\). Instead, above \(q=2000\) ridge begins to provide better results. For long time series, at \(q=20,000\), the best performing model is ridge with parameters \(\gamma =10^{-5}\), p-value \(p_v = 5\;10^{-6}\). LoGo is also performing well when time series are long with best performance obtained at \(q=20,000\) for parameters \(\gamma =10^{-10}\), p-value \(p_v = 5\;10^{-6}\). We note that LoGo instead performs poorly in the region of parameters with \(\gamma \le 0.1\) and \(p_v \le 0.01\) for short time-series \(q\le p/2\).
# Conclusions and perspectives {#s.conclusions}
In this paper we have undertaken the challenging task to explore models and parameter regions where the analytics of time series can retrieve significant fractions of true causality links from linear multivariate autoregressive process with known causality structure. Results demonstrate that sparse models with conditional transfer entropy are the ones who achieve best results with significant causality link retrievals already for very short time series even with \(q \le p/5 = 20\). Unexpectedly, we observed that the structure of the inference networks in the two sparse models, Glasso and LoGo, have excessive numbers of false positives yielding to rather poor performances. However, in these models false positive can be efficiently filtered out by imposing statistical significance of the transfer entropies. Results are affected by the choice of the parameters and the fact that the models depend on various parameters (\(q\), \(p\), \(\gamma\), \(p_v\), \(P\)) make the navigation in this space quite complex. We observed that the choice of p-values, \(p_v\), for valid transfer entropies affects results. Within our setting we obtained best results with the smaller p-values especially in the regions of short time-series. We note that the regularizer parameter \(\gamma\) also plays an important role and best performances are obtained by combination of the two parameters \(\gamma\) and \(p_v\). Not surprizingly, longer time-series yield to better results. We observe that conditioning to all other variables or unconditioning is affecting the transfer entropy estimation with better performing causality network retrieval obtained for conditioned transfer entropies. However, qualitatively, results are comparable. Other intermediate cases, such as conditioning to past of all other variables only, have been explored again with qualitatively comparable results. It must be said that in the present system results are expected to be robust to different conditionings because the underlying network of the investigated processes is sparse. For denser inference structures, conditioning could affect more the results. Consistently with the findings in we find that LoGo outperforms the other methods. This is encouraging because the present settings of LoGo is using a simple class of information filtering networks, namely the TMFG, obtained by retaining largest correlations. There are a number of alternative information filtering networks which should be explored. In particular, given the importance of statistical validation emerged from the present work, it would be interesting to explore statistical validation within the process of construction of the information filtering networks themselves. In this paper we investigate a simple case with a linear autoregressive multivariate normal process analysed by means of linear models. Both LoGo and Glasso can be extended to the non-linear case with LoGo being particularly suitable for non-parametric approaches as well. There are Alternative methods to extract causality networks from short time series, in particular Multispatial CCM appears to perform well for short time series. A comparison between different approaches and the application of these methods to real data will be extremely interesting. However this should be the object of future works. | {'timestamp': '2017-10-06T02:07:21', 'yymm': '1706', 'arxiv_id': '1706.01954', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01954'} |
null | null |
null | null |
null | null |
# INTRODUCTION {#sec1}
Recent observations in the \(\gamma\)-ray A key feature of the non-thermal emission from blazars is the distinct variability at all frequencies (with different variability time scales-from years down to a few minutes). The shortest variability time scales are usually observed for the highest energy band; an example is the minute scale variability of PKS 2155-304 and IC 310 which implies that the emission is produced in a very compact region. Therefore, by observing blazars one gets a unique chance to investigate the jet structure on sub-parsec scales.\
By their emission line features blazars are commonly grouped as BL Lacertae objects (BL-Lacs) and Flat-Spectrum Radio Quasars (FSRQs). BL Lacs have weak or no emission lines, while FSRQs have stronger emission lines. The difference in the emission-line properties of FSRQs and BL Lacs may be connected with that in the properties of accretion in these objects.\
The multiwavelength observations of blazars have shown that their Spectral Energy Distribution (SED) has two broad non-thermal peaks-one at the IR/optical/UV/X-ray and the other at the Higher-Energy (HE; \(>\) 100 MeV) \(\gamma\)-ray The majority of the blazars detected in VHE \(\gamma\)-ray Among FSRQs, PKS 1441+25 is one of the most distant sources detected so far at z=0.939. In April 2015 both VERITAS and MAGIC collaborations announced the detection of VHE \(\gamma\)-rays In the theoretical interpretation of the multiwavelength emission from blazars, the size/location of the emitting region, magnetic field and electron energy distribution are uncertain. Only during flaring periods some of the unknown parameters can be constrained based on the observations in different bands. The observations of PKS 1441+25 during the bright period in April 2015 by different instruments provide us with data on the maximums of the emitting components (Swift UVOT/ ASAS-SN and *Fermi* LAT) as well as on the transition region between these components in the energy range from 0.3 to 30 keV (Swift XRT and NuSTAR). Similar data (up to HE \(\gamma\)-ray This paper is structured as follows. The results of the spectral and temporal analysis of the *Fermi* LAT data are presented in Section [2](#sec2){reference-type="ref" reference="sec2"}. The broadband SED modeling with MCMC technique is presented in Section [3](#sec3){reference-type="ref" reference="sec3"} and discussion and conclusions are presented in Section [4](#sec4){reference-type="ref" reference="sec4"}.
# *Fermi* LAT DATA ANALYSIS {#sec2}
The large Area Telescope on board the *Fermi* satellite is a pair-conversion telescope sensitive to \(\gamma\)-rays In the present paper, for spectral analysis we use the publicly available data acquired in the periods from January 06 to 28 and from April 15 to 26, 2015. These two periods have been picked, because they are contemporaneous with the Swift XRT observations of the source. The data were analyzed with the standard *Fermi* Science Tools v10r0p5 software package released on May 18, 2015 available from the *Fermi* Science Support Center [^1]. The latest reprocessed PASS 8 events and spacecraft data are used with the instrument response function P8R2\_ SOURCE\_ V6. We have downloaded photons in the energy range from 100 MeV to 100 GeV from a region of interest defined as a circle of a \(20^{\circ}\) radius centered at the \(\gamma\)-ray
## Spectral analysis {#sec2.2}
In order to find the best matches between spectral models and events, an unbinned likelihood analysis is performed with *gtlike*. The PKS 1441+25 spectrum has been initially modeled as a power-law function where the normalization and the power-law index are taken as free parameters. The best fit parameters obtained with *gtlike* analysis are presented in Table [1](#table:param){reference-type="ref" reference="table:param"} and the corresponding spectrum is shown in Fig. [\[fg1\]](#fg1){reference-type="ref" reference="fg1"} (blue and red data for January and April, respectively). The spectrum is calculated by separately running *gtlike* for 5 energy bands equal on a log scale.\
The fluxes presented in Table [1](#table:param){reference-type="ref" reference="table:param"} significantly exceed the averaged flux given in 3FGL (\(\approx1.28\times10^{-8}\:{\rm photon\:cm^{-2}\:s^{-1}}\)). The photon index estimated in January 2015 is consistent with the value reported in 3FGL \(\alpha=2.13\) (averaged over 4 years of observations); however, a relative hardening of \(\alpha=1.74\pm0.06\) is observed in April, which is rarely observed for FSRQs. Moreover, we note an indication of deviation of the power-law model with respect to the data above 10s of GeV energies observed in April (red bowtie plot in Fig. [\[fg1\]](#fg1){reference-type="ref" reference="fg1"}). In order to check for a statistically significant curvature in the spectrum, an alternative fit of the power-law with an exponential cutoff function in the form of \(dN/dE\sim E_{\gamma}^{-\alpha}\:\times Exp(-E_{\gamma}/E_{cut})\) is done, which results in \(\alpha=1.56\pm0.1\) and \(E_{\rm cut}=17.7\pm8.9\) GeV (black bowtie plot in Fig. [\[fg1\]](#fg1){reference-type="ref" reference="fg1"}). The power-law and cutoff models are compared with a log likelihood ratio test: the TS is twice the difference in the log likelihoods, which gives 8 for this case. Note that the TS probability distribution can be approximated by a \(\chi^2\) distribution with 1 degree of freedom (dof) corresponding to the difference of the dof between the two functions. The results give \(P(\chi^2)=0.0046\), which again indicates a deviation from a simple power-law function. The best-fit cutoff power-law function is shown as a black bowtie line in Fig. [\[fg1\]](#fg1){reference-type="ref" reference="fg1"}. However, \(2.8\:\sigma\) is not a high enough significance to claim for a statistically significant curvature although it is as high as \(3.86\:\sigma\) if the data collected during the whole month of April are considered.
## Temporal analysis {#sec2.3}
In order to investigate the size of the \(\gamma\)-ray The light curve of PKS 1441+25 for the period from January to December 2015 has been calculated by the *gtlike* tool, applying the unbinned likelihood analysis method. \((0.1-100)\) GeV photons from a region with a \(10^{\circ}\) radius centered on the position of PKS 1441+25 are used in the analysis with the appropriate quality cuts applied as in the previous case. During the analysis, in order to reduce the uncertainty in the flux estimations, in the model file the photon indices of all background sources are fixed to the best guess values. Two different sets of light curves are calculated, considering the power-law index of PKS 1441+25 as being fixed and then as free. Since no variability is expected for the background diffuse emission, the normalization of both background components is also fixed to the values obtained for the whole time period.\
The \(\gamma\)-ray Next, in order to investigate the flux changes in time, and in particular in the flaring periods, the light curves have been generated by an adaptive binning method. In this method, the time bin widths are flexible and chosen to produce bins with constant flux uncertainty. This method allows detailed investigation of the flaring periods, since at times of a high source flux, the time bins are narrower than during lower flux levels, therefore the rapid changes of the flux can be found. In order to reach the necessary relative flux uncertainty, the integral fluxes are computed above the optimal energies which correspond to \(E_0\) = 215.4 MeV in this case.\
Adaptively binned light curves in the 215 MeV-300 GeV energy range with 20% and 15% uncertainties have been generated. Flare is present in both light curves. The light curve with 20% flux uncertainty at each bin is presented in Fig. [\[fg2\]](#fg2){reference-type="ref" reference="fg2"} (b) with red color. It confirms all the features visible in the constant-bin-width light curve, but also allows us to investigate fast variability during high-flux states in greater detail. The first flare episode occurred during MJD 57043.30-57049.38, when the time width was less than \(\sim15\) hours. A strong flaring period is observed around the 24th-25th of January. The flux peak of \((1.14\pm0.24)\times10^{-6}\;{\rm photon\:cm^{-2}\:s^{-1}}\) was observed on the 24th of January at 22:35 \(pm\) in a bin with a half-width of 3.1 hours. The analysis of the data acquired in the mentioned period on the energies of \(>100\) MeV results in a flux of \((2.22\pm0.38)\times10^{-6}\;{\rm photon\:cm^{-2}\:s^{-1}}\), which is the highest photon flux detected from this source. The data analysis for the entire flaring period (January 21-28) resulted in a flux of \((1.05\pm0.06)\times10^{-6}\;{\rm photon\:cm^{-2}\:s^{-1}}\) and a photon index of \(\sim1.98\pm0.04\). After MJD 57049.38, PKS 1441+25 was in its quiescent state, and the data should be accumulated for more than a day to reach 20 % uncertainty. Then from MJD 57109.89 to MJD 57143.91, PKS 1441+25 was again in its active state which was characterized by emission with a significantly hardened \(\gamma\)-ray Furthermore, to derive the flare doubling timescales and understand the nature of the January flare, the light curve is fitted with an exponential function in the form of \[F(t)= F_{\rm c}+F_0\times\left(e^{\frac{t-t_{\rm 0}}{t_{\rm r}}}+e^{\frac{t_{\rm 0}-t}{t_{\rm d}}}\right)^{-1} \label{func}\] where \(t_{0}\) is the time of the maximum intensity of the flare (\(F_0\)), \(F_c\) is the constant level present in the flare, \(t_{\rm r}\) and \(t_{\rm d}\) are the rise and decay time constants, respectively. The fit shows that the flare is best explained when \(t_{0}=57048.25\pm0.18\), \(t_r= 1.92 \pm 0.3\), \(t_d= 0.72 \pm 0.1\) and \(F_0=(22.6 \pm 1.4)\times 10^{-7}\mathrm{photon\:cm^{-2}s^{-1}}\). The fit of the flaring period is shown in Fig. [\[fg2_fit\]](#fg2_fit){reference-type="ref" reference="fg2_fit"} with a dashed red line. Using this technique, it is also possible to estimate the shortest time variability (flux doubling) defined by \(\tau=2 t_{\rm r,d}\), corresponding to \(\tau=1.44\) days which is used to put an important constraint on the radiative region size. We note that the previous PKS 1441+25 \(\gamma\)-ray
# BROADBAND SED MODELLING {#sec3}
It is hard to make theoretical modeling of the observed broadband SED because the structure of the central region of blazars is complex and the exact localization of emitting regions is unknown. The observed fast variability indicates compactness of the emitting region but its localization remains an open problem. Along the jet, the emission can be produced in different zones and depending on the distance from the central black hole different components can contribute to the observed emission.
## Broadband SED
The broadband SEDs of PKS 1441+25 for different periods are shown in Fig. [\[sed\]](#sed){reference-type="ref" reference="sed"} where with red and blue colors are the SED observed in January and April respectively, while the archival data from ASI science data center [^2] are shown with gray color. We note that during the high states, the second emission peak increased by intensity and shifted to HEs. This kind of change has already been observed during the flaring state of 4C+49.22 and PKS 1510-089. During the flaring periods the low-energy component's intensity increased as compared with the quiescent state; the increase in April exceeded that one observed in January (although the power-law photon index in the X-ray band (\(\approx 2.3\)) had been relatively constant during both observations). More evident and drastic is the change of the peak intensity of the low energy component; from January to April it increased by nearly an order of magnitude and as compared with the quiescent state it increased \(\geq 15\) times. On the contrary, the peak of the second component (in the HE \(\gamma\)-ray Such a strong amplification of the emission from blazars can be explained by means of introducing changes in the emission region parameters, e.g., in the magnetic field, emitting region size, bulk Lorentz factor and others, and/or particle energy distribution. In principle, all the parameters describing the emitting region can be changed at the same time if the flares are due to a global change in the physical processes in the jet, which also affect the jet dynamics and properties. However, usually, the change in one or two parameters is enough to explain the flares. An interesting study of the flaring activity in FSRQs as a result of changes in different parameters has been investigated in. Namely, the emission spectra evolution as a function of changes in different parameters (e.g., bulk Lorentz factor, magnetic field, accretion rate, etc.) is investigated. In the case of PKS 1441+25, during its flaring periods, both the low energy and HE components increased several times. The increase of the second component is most likely due to moving of the emitting region outside its BLR. In principle, there are two possibilities: *i)* either the emitting region moves faster due to increasing bulk Lorentz factor and leaves the BLR or ii) the bulk Lorentz factor is unchanged and only the emitting region is moving beyond the BLR. In the first case, since the external photon density in the comoving frame of the jet depends on the Doppler boosting factor, a strong increase in the Compton dominance will be observed. We note that the change of the bulk Lorentz factor will also affect the low energy component. In the second case, the flaring activity is due to the change of the location of the emitting region and due to the magnetic field amplification. Additional increase of the magnetic field from January to April is also evident when the low energy component kept increasing (this corresponds to the case shown in Fig. 1 (b) in ). Accordingly, we discuss two possibilities. First, we assume that \(\delta\) has increased from 10 in the quiescent to 18 in the flaring periods, and then we assume that it was constant (\(\delta=18\)) in both periods. These values are below and above the estimated mean bulk Lorentz factor of FSRQs obtained from the analysis of a large sample of \(\gamma\)-ray
## Theoretical modeling
We attempt to fit the SEDs in the high states of January and April as well as in the quiescent state. Even if a quiescent state SED is constrained with non-simultaneous data, its modeling provides an insight into the dominant physical processes which are constantly present in the jet but are covered by the flaring components during the high states. We modeled the PKS 1441+25 SED for high and quiescent states in the framework of single-zone leptonic models that include the synchrotron, SSC, and EC processes. The emission region (the \"blob\"), assumed to be a sphere with a radius of \(R\) which is moving with a bulk Lorentz factor of \(\Gamma\), carries a magnetic field with an intensity of \(B\) and a population of relativistic electrons. The blob velocity makes a small angle with respect to the line of sight, so the emission is amplified by a relativistic Doppler factor of \(\delta\). The energy spectrum of the population of electrons in the jet frame, which is responsible for the non-thermal emission is assumed to have a broken power-law shape: \[N_e^{\prime}(E_{e}^\prime)= \begin{cases} N_0^\prime \left(\frac{E_{e}^\prime}{m_ec^2}\right)^{-\alpha_1} & E_\mathrm{e,min}^\prime \leqslant E_{e}^\prime \leqslant E_\mathrm{br}^\prime \\ N_0^\prime \left(\frac{E_{br}^\prime}{m_ec^2}\right)^{\alpha_2-\alpha_1} \left(\frac{E_{e}^\prime}{m_ec^2}\right)^{-\alpha_2} & E_\mathrm{br}^\prime \leqslant E_{e}^\prime \leqslant E_{e,max}^\prime \end{cases} \label{BPL}\] where \(N_{0}^\prime\) is connected with the total electron energy \(U_{\rm e}=\int_{\rm E_{min}^\prime}^{\rm E_{\rm max}^\prime}E_{e}^\prime N_e(E_{e}^\prime)dE_e^\prime\), \(\alpha_{1}\) and \(\alpha_{2}\) are the low and high indexes of electrons correspondingly below and above the break energy \(E_{\rm br}^\prime\), and \(E_{\rm min}^\prime\) and \(E_{\rm max}^\prime\) are the minimum and maximum energies of electrons in the jet frame, respectively. The electron spectrum given in Eq. [\[BPL\]](#BPL){reference-type="ref" reference="BPL"} is naturally formed from the cooling of relativistic electrons.\
The low-energy (from radio to optical/X-ray) emission is due to the synchrotron emission of electrons with an energy spectrum as given by Eq. [\[BPL\]](#BPL){reference-type="ref" reference="BPL"} in a homogeneous and randomly oriented magnetic field. For the quiescent state we assume the energy dissipation occurs close to the central source region and it is explained as an IC scattering of synchrotron photons (SSC). Instead the high state emission is dominated by that from a region well outside the BLR in order to avoid the strong absorption of photons with energies \(\geq\:100\:{\rm GeV}\) (similar assumptions have been already made in ). In this case the dominant external photon field is the IR radiation from the dusty torus which, as we assume, has a blackbody spectrum with a luminosity of \(L_{\it IR}=\eta\:L_{\it disc}\) (\(\eta=0.6\), ) and a temperature of \(T=10^{3}\) K and fills a volume that for simplicity is approximated as a spherical shell with a radius of \(R_{IR}=3.54\times10^{18}\:(L_{\rm disc}/10^{45})^{0.5}\) cm. The disc luminosity is estimated from the BLR luminosity, \(L_{\rm disc}=10\times L_{\rm BLR}\approx 2\times10^{45}{\rm erg\:s^{-1}}\).
## Fitting technique
In order to constrain the model parameters more efficiently, we employed the MCMC method, which enables to derive the confidence intervals for each model parameter. For the current study we have modified the *naima* package which derives the best-fit and uncertainty distributions of spectral model parameters through MCMC sampling of their likelihood distributions. The prior likelihood, our prior knowledge of the probability distribution of a given model parameter and the data likelihood functions are passed onto the emcee sampler function for an affine-invariant MCMC run. In addition, there are multiple simultaneous walkers which improve the efficiency of the sampling and reduce the number of computationally expensive likelihood calls. We run the sampling with 64 simultaneous walkers, for 100 steps of burn-in, and 100 steps of run. In the parameter sampling, the following expected ranges are considered: \(1.5\leq(\alpha_{1,2})\leq10\), \(0.511\:{\rm MeV}\leq E^\prime_{(br,\:min,\:max)}\leq1\:{\rm TeV}\), and \(N_0\) and \(B\) are defined as positive parameters. The synchrotron emission is calculated using the parameterization of the emissivity function of synchrotron radiation in random magnetic fields presented in while the IC emission is computed based on the monochromatic differential cross section of.
:::
## SED modeling and results
The results of SED modeling are shown in Fig. [\[sed\]](#sed){reference-type="ref" reference="sed"} with the corresponding parameters in Table [\[table_fit\]](#table_fit){reference-type="ref" reference="table_fit"}. The radio emission is due to the low-energy electrons which are accumulated for longer periods, that is why, the radio data are treated as an upper limit for the purposes of our modeling. To have an indication of a change in the energetic contents of the jet, as well as of changes in the radiating particle distribution, first we try to fit the SED in a quiescent state which is modeled assuming two different Doppler boosting factors. The gray solid line in Fig. [\[sed\]](#sed){reference-type="ref" reference="sed"} shows the synchrotron/SSC emission assuming that the jet Doppler boosting factor is \(\delta=10\), and the gray dashed line is the case of \(\delta=18\). In case of \(\delta=10\), as the emitting region size is as small as \(R_b=1.92\times10^{16}\) cm, the magnetic field should be as strong as \(B=0.19\) G to account for the observed data, while at \(\delta=18\) the magnetic field is much weaker, \(B=0.046\) G. Also, the underlying electron distribution for the case of \(\delta=10\) is characterized by a slightly higher break (\(2.83\) GeV versus \(1.11\) GeV) in order to account for the observed emission.\
The emission in flaring periods is modeled assuming that the HE emission is entirely due to the IC scattering of external photons (Fig. [\[sed\]](#sed){reference-type="ref" reference="sed"}). In all calculations the absorption due to the EBL was taken into account using a model from since a strong absorption is evident at \(\geq\:100\) GeV (red dashed line in Fig. [\[sed\]](#sed){reference-type="ref" reference="sed"}). In both periods the HE electron spectral index is within the range of \(\alpha_2\sim (3.46-3.64)\) which is required to explain the UV-X-ray data with a photon index of \(\approx2.3\). The lack of low-energy data makes the precise estimation of the low energy electron index harder. Only the Swift XRT/NuSTAR data from the observation of the transition region between low and high energy components allows to define the parameters \(E_{\rm min}\) and \(\alpha_1\). The low energy electron index is in a typical range expected from shock acceleration theories, \(\alpha_1\approx2\).\
As distinct from the quiescent state, in order to explain the flaring activities, both, the electron distribution and the magnetic field should be varying. We note that the magnetic field required for modeling of flaring periods, (\(B\geq\) 0.11 G), is weaker than that one estimated in the quiescent state in case of \(\delta=10\) (\(B\sim 0.19\) G). Since the synchrotron emission depends on the total number of emitting electrons \(N_{\rm e}\), \(\delta\) and magnetic field strength \(B\), in case of smaller \(\delta\) (and emitting region size) the magnetic field should be stronger. Instead, when \(\delta\) is constantly equal to 18 in both states, the magnetic field should be nearly \(\sim\) 2.4 and \(\sim\) 3.9 times stronger in January and April, respectively, in order to explain the observed data. As the synchrotron photon density is proportional to \(B^2\), the increase in the magnetic field strength resulted in the observed increase of the synchrotron flux by a factor of \(5.7\) and \(15.3\) (Fig. [\[sed\]](#sed){reference-type="ref" reference="sed"}). In the modeling of the SEDs observed in the flaring periods of January and April, the magnetic field should be changed in accordance with the increase in the low energy component. Since the emission in the HE band is dominated by the IC scattering of external photons, this component remains stable during those periods (this corresponds to Fig. 1 (b) ).\
The electron spectra obtained during the fit of SEDs in quiescent and flaring states are shown in Fig. [\[elect\]](#elect){reference-type="ref" reference="elect"}. It is clear the evolution of the electron spectra during the quiescent and high states. The low energy indexes of the underlying electron distribution in the quiescent state (\(\delta=10\)) are softer as compared with the flaring period (April). The total electron energy for modeling the emission in the quiescent period, when \(\delta=18\), is almost of the same order as that in the flaring periods, which is expected, as the magnetic field is weaker, most of the jet energy is carried by particles. During the flaring periods, there are evident changes also in the underlying electron distribution. The electron distribution best describing the data observed in April hints at *i)* hardening of the low energy index, *ii)* a higher break at \(\sim 3.1\) GeV and maximum energies of \(\sim203\) GeV. \(E_{\rm br}\) and \(E_{\rm max}\) are expected to shift, as the \(\gamma\)-ray Similar modeling of the SED of PKS 1441+25 observed in April has been already done in, but it was done in a different manner. For example, in the low electron energy index is fixed to be \(\alpha_1=2\), a value expected from strong shock acceleration theories, while in our case all the parameters can vary in the fitting procedure. After having observed the hardening of the \(\gamma\)-ray
# DISCUSSION and CONCLUSIONS {#sec4}
In this paper, we present the results of the study of the \(\gamma\)-ray The \(\gamma\)-ray After the flare on MJD 57049.38, the source is in its quiet state and the next increase in the flux is observed starting from MJD 57109.89. Even if during this period, the flux amplitude is lower than one that observed in January, an interesting modification of the \(\gamma\)-ray
##### The origin of multiwavelength emission:
The SEDs observed during quiescent and flaring states are modeled using one-zone leptonic models and the model parameters are estimated using the MCMC method. The HE \(\gamma\)-ray The flares observed in January and April can be explained assuming there are changes in the bulk Lorentz factor or in the magnetic field. If the emitting region leaves the BLR region due to the increase of the bulk Lorentz factor (from \(\delta=10\) to \(\delta=18\)), the Compton dominance will increase as it has been observed in the \(\gamma\)-ray When comparing the electron parameters required for the modeling of the SEDs in January and April, we find a hint of possible hardening of the low energy electron index in April. We note, however, that no definite conclusions can be drawn since \(\alpha_1\) is poorly constrained (due to missing or nonsufficient data). For all that, the April hardening of the \(\gamma\)-ray
##### Jet Energetics:
The jet power in the form of magnetic field and electron kinetic energy are calculated by \(L_{B}=\pi c R_b^2 \Gamma^2 U_{B}\) and \(L_{e}=\pi c R_b^2 \Gamma^2 U_{e}\), respectively, and are given in Table [\[table_fit\]](#table_fit){reference-type="ref" reference="table_fit"}. The jet power in the electrons changes in the range \((4.5-9.6)\times10^{45}\:{\rm erg\:s^{-1}}\) during the flares, while in the quiescent state it is of the order of \((2.1-4.1)\times10^{45}\:{\rm erg\:s^{-1}}\). Assuming one proton per relativistic electron (e.g., ), the total kinetic energy in the jet is \(L_{\rm kin}=8.02\times10^{47}\:{\rm erg\:s^{-1}}\) and \(L_{\rm kin}=1.35\times10^{47}\:{\rm erg\:s^{-1}}\) for January and April, respectively.\
The maximum \(\gamma\)-ray\
The observations in both X-ray and \(\gamma\)-ray The multiwavelength observations of PKS 1441+25 during the flaring periods allowed us to investigate and discuss the changes that possibly took place in the jets and caused flaring activities. However, the parameters describing the underlying electron distribution below the break are poorly constrained, because the data describing the rising part of both low and HE components are missing. It did not allow us to exactly identify the processes responsible for the acceleration of particles in the jet. However, the future possible observations of flaring periods also in other energy bands will provide a chance to investigate the dominant particle acceleration processes. | {'timestamp': '2017-06-07T02:03:27', 'yymm': '1706', 'arxiv_id': '1706.01611', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01611'} |
# Introduction
Recently, character composition models have shown great success in many NLP tasks, mainly because of their robustness in dealing with out-of-vocabulary (OOV) words by capturing sub-word informations. Among the character composition models, bidirectional long short-term memory (LSTM) models and convolutional neural networks (CNN) are widely applied in many tasks, e.g. part-of-speech (POS) tagging, named entity recognition, language modeling, machine translation and dependency parsing. In this paper, we present a state-of-the-art general-purpose tagger that uses CNNs both to compose word representations from characters and to encode context information for tagging.[^1] We show that the CNN model is more capable than the LSTM model for both functions, and more stable for unseen or unnormalized words, which is the main benefit of character composition models. compared the performance of CNN and LSTM as character composition model for dependency parsing, and concluded that CNN performs better than LSTM. In this paper, we show that this is also the case for POS tagging. Furthermore, we extend the scope to morphological tagging and supertagging, in which the tag set is much larger and long-distance dependencies between words are more important. In these three tagging tasks, we compare our tagger with the **bilstm-aux** tagger and the CRF-based morphological tagger **MarMot**. The CNN tagger shows robust performance accross the three tasks, and achieves the highest average accuracy in all tasks. It (significantly) outperforms LSTM in morphological tagging, and outperforms both baselines in supertagging by a large margin. To test the robustness of the taggers against the OOV problem, we also conduct experiments using artificially constructed unnormalized text by corrupting words in the normal dev set. Again, the CNN tagger outperforms the two baselines by a very large margin. Therefore we conclude that our CNN tagger is a robust state-of-the-art general-purpose tagger that can effectively compose word representation from characters and encode context information.
# Model
Our proposed CNN tagger has two main components: the character composition model and the context encoding model. Both components are essentially CNN models, capturing different levels of information: the first CNN captures morphological information from character n-grams, the second one captures contextual information from word n-grams. Figure [\[fig:cnn\]](#fig:cnn){reference-type="ref" reference="fig:cnn"} shows a diagram of both models of the tagger.
## Character Composition Model
The character composition model is similar to , where several convolution filters are used to capture character n-grams of different sizes. The outputs of each convolution filter are fed through a max pooling layer, and the pooling outputs are concatenated to represent the word.
## Context Encoding Model
The context encoding model captures the context information of the target word by scanning through the word representations of its context window. The word representation could be only word embeddings (\(\vec{w}\)), only composed vectors (\(\vec{c}\)) or the concatenation of both (\(\vec{w}+\vec{c}\)) A context window consists of N words to both sides of the target word and the target word itself. To indicate the target word, we concatenate a binary feature to each of the word representations with 1 indicating the target and 0 otherwise, similar to . Additional to the binary feature, we also concatenate a position embedding to encode the relative position of each context word, similar to .
## Hyper-parameters {#hyper}
For the character composition model, we take a fixed input size of 32 characters for each word, with padding on both sides or cutting from the middle if needed. We apply four convolution filters with sizes of 3, 5, 7, and 9. Each filter has an output channel of 25 dimensions, thus the composed vector is 100-dimensional. We apply Gaussian noise with standard deviation of 0.1 is applied on the composed vector during training. For the context encoding model, we take a context window of 15 (7 words to both sides of the target word) as input and predict the tag of the target word. We also apply four convolution filters with sizes of 2, 3, 4 and 5, each filter is stacked by another filter with the same size, and the output has 128 dimensions, thus the context representation is 512-dimensional. We apply one 512-dimensional hidden layer with ReLU non-linearity before the prediction layer. We apply dropout with probability of 0.1 after the hidden layer during training. The model is trained with averaged stochastic gradient descent with a learning rate of 0.1, momentum of 0.9 and mini-batch size of 100. We apply L2 regularization with a rate of \(10^{-5}\) on all the parameters of the network except the embeddings.
# Experiments
## Data
We use treebanks from version 1.2 of Universal Dependencies[^2] (UD), and in the case of several treebanks for one language, we only use the canonical treebank. There are in total 22 treebanks, as in .[^3] Each treebank splits into train, dev, and test sets, we use the dev sets for early stop, and test on the test sets.
## Tasks
We evaluate our method on three tagging tasks: POS tagging (**[Pos]{.smallcaps}**), morphological tagging (**[Morph]{.smallcaps}**) and supertagging (**[Stag]{.smallcaps}**). For POS tagging we use Universal POS tags, which is an extension of . The universal tag set tries to capture the "universal" properties of words and facilitate cross-lingual learning. Therefore the tag set is very coarse and leaves out most of the language-specific properties to morphological features. Morphological tags encode the language-specific morphological features of the words, e.g., number, gender, case. They are represented in the UD treebanks as one string which contains several key-value pairs of morphological features.[^4] Supertags are tags that encode more syntactic information than standard POS tags, e.g. the head direction or the subcategorization frame. We use dependency-based supertags which are extracted from the dependency treebanks. Adding such tags into feature models of statistical dependency parsers significantly improves their performance. Supertags can be designed with different levels of granularity. We use the standard Model 1 from , where each tag consists of head direction, dependency label and dependent direction. Even with the basic supertag model, the [Stag]{.smallcaps} task is more difficult than [Pos]{.smallcaps} and [Morph]{.smallcaps} because it generally requires taking long-distance dependencies between words into consideration. We select these tasks as examples for tagging applications because they differ strongly in tag set sizes. Generally, the [Pos]{.smallcaps} set sizes for all the languages are no more than 17 and [Stag]{.smallcaps} set sizes are around 200. When treating morphological features as a string (i.e. not splitting into key-value pairs), the sizes of the [Morph]{.smallcaps} tag sets range from about 100 up to 2000.
## Setups
As baselines to our models, we take the two state-of-the-art taggers MarMot[^5] (denoted as CRF) and bilstm-aux[^6] (denoted as LSTM). We train the taggers with the recommended hyper-parameters from the documentation. To ensure a fair comparison (especially between LSTM and CNN), we generally treat the three tasks equally, and do not apply task-specific tuning on them, i.e., using the same features and same model hyper-parameters in each single task. Also, we do not use any pre-trained word embeddings. For the LSTM tagger, we use the recommended hyper-parameters in the documentation[^7] including 64-dimensional word embeddings (\(\vec{w}\)) and 100-dimensional composed vectors (\(\vec{c}\)). We train the \(\vec{w}\), \(\vec{c}\) and \(\vec{w}+\vec{c}\) models as in . We train the CNN taggers with the same dimensionalities for word representations. For the CRF tagger, we predict [Pos]{.smallcaps} and [Morph]{.smallcaps} jointly as in the standard setting for MarMot, which performs much better than with separate predictions, as shown in and in our preliminary experiments. Also, it splits the morphological tags into key-value pairs, whereas the neural taggers treat the whole string as a tag.[^8] We predict [Stag]{.smallcaps} as a separate task.
## Results
The test results for the three tasks are shown in Table [\[tbl:results\]](#tbl:results){reference-type="ref" reference="tbl:results"} in three groups. The first group of seven columns are the results for [Pos]{.smallcaps}, where both LSTM and CNN have three variations of input features: word only (\(\vec{w}\)), character only (\(\vec{c}\)) and both (\(\vec{w}+\vec{c}\)). For [Morph]{.smallcaps} and [Stag]{.smallcaps}, we only use the \(\vec{w}+\vec{c}\) setting for both LSTM and CNN. On macro-average, three taggers perform close in the [Pos]{.smallcaps} task, with the CNN tagger being slightly better. In the [Morph]{.smallcaps} task, CNN is again slightly ahead of CRF, while LSTM is about 2 points behind. In the [Stag]{.smallcaps} task, CNN outperforms both taggers by a large margin: 2 points higher than LSTM and 8 points higher than CRF. While considering the input features of the LSTM and CNN taggers, both taggers perform close with only \(\vec{w}\) as input, which suggests that the two taggers are comparable in encoding context for tagging [Pos]{.smallcaps}. However, with only \(\vec{c}\), CNN performs much better than LSTM (95.54 vs. 92.61), and close to \(\vec{w}+\vec{c}\) (96.18). Also, \(\vec{c}\) consistently outperforms \(\vec{w}\) for all languages. This suggests that the CNN model alone is capable of learning most of the information that the word-level model can learn, while the LSTM model is not. The more interesting cases are [Morph]{.smallcaps} and [Stag]{.smallcaps}, where CNN performs much higher than LSTM. We hypothesize three possible reasons to explain the considerably large difference. First, the LSTM tagger may be more sensitive to hyper-parameters and requires task specific tuning. We use the same setting which is tuned for the [Pos]{.smallcaps} task, thus it underperforms in the other tasks. Second, the LSTM tagger may not deal well with large tag sets. The tag set size for [Morph]{.smallcaps} are larger than [Pos]{.smallcaps} in orders of magnitudes, especially for Czech, Basque, Finnish and Slovene, all of which have more than 1000 distinct [Morph]{.smallcaps} tags in the training data, and the LSTM performs poorly on these languages. Third, the LSTM has theoretically unlimited access to all the tokens in the sentence, but in practice it might not learn the context as good as the CNN. In the LSTM model, the information of long-distance contexts will gradually fade away during the recurrence, whereas in the CNN model, all words are treated equally as long as they are in the context window. Therefore the LSTM underperforms in the [Stag]{.smallcaps} task, where the information from long-distance context is more important.
:::
## Unnormalized Text
It is a common scenario to use a model trained with news data to process text from social media, which could include intentional or unintentional misspellings. Unfortunately, we do not have social media data to test the taggers. However, we design an experiment to simulate unnormalized text, by systematically editing the words in the dev sets with three operations: insertion, deletion and substitution. For example, if we modify a word *abcdef* at position 2 (0-based), the modified words would be *abxcdef*, *abdef*, and *abxdef*, where *x* is a random character from the alphabet of the language. For each operation, we create a group of modified dev sets, where all words longer than two characters are edited by the operation with a probability of 0.25, 0.5, 0.75, or 1. For each language, we use the models trained on the normal training sets and predict [Pos]{.smallcaps} for the three groups of modified dev set. The average accuracies are shown in Figure [\[fig:edits\]](#fig:edits){reference-type="ref" reference="fig:edits"}. Generally, all models suffer from the increasing degrees of unnormalized texts, but CNN always suffers the least. In the extreme case where almost all words are unnormalized, CNN performs 4 to 8 points higher than LSTM and 4 to 11 points higher than CRF. This suggests that the CNN is more robust to misspelt words. While looking into the specific cases of misspelling, CNN is more sensitive to insertion and deletion, while CRF and LSTM are more sensitive to substitution.
# Conclusion
In this paper, we propose a general-purpose tagger that uses two CNNs for both character composition and context encoding. On the universal dependency treebanks (v1.2), the tagger achieves state-of-the-art results for POS tagging and morphological tagging, and to the best of our knowledge, it also performs best for supertagging. The tagger works well across different tagging tasks without tuning the hyper-parameters, and it is also robust against unnormalized text.
[^1]: We will release the code in the camera-ready version.
[^2]: <http://universaldependencies.org>
[^3]: We use all training data for Czech, while only use a subset.
[^4]: German, French and Indonesian do not have [Morph]{.smallcaps} tags in UD-1.2, thus not evaluated in this task.
[^5]: <http://cistern.cis.lmu.de/marmot/>
[^6]: <https://github.com/bplank/bilstm-aux>
[^7]: We use the most recent version of the tagger and stacking 3 layers of LSTM as recommended. The average accuracy for [Pos]{.smallcaps} in our evaluation is slightly lower than reported in the paper, presumably because of different versions of the tagger, but it does not influence the conclusion.
[^8]: Since we use the CRF tagger as a non-neural baseline model, we prefer to use the settings which maximize its performances than the rigorously equal but suboptimal settings. | {'timestamp': '2017-06-07T02:05:46', 'yymm': '1706', 'arxiv_id': '1706.01723', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01723'} |
null | null |
# Introduction
Flows in suspensions are often characterized by transport phenomena that originate from the interaction of the particles among themselves and with the flow. In dilute suspensions, the particle interactions are mediated by hydrodynamic forces. In more concentrated suspensions, the contribution from direct contact of the particles is dominant. Effective diffusivities greatly exceeding the values expected from molecular effects are often observed, even in the absence of turbulence . A medium of particular interest in the geophysical context is grease ice, that is the dense slurry of ice crystals often observed at the water surface in polar seas during periods of strong wind and freezing temperatures. Consolidation of the slurry to pack ice rests on efficient removal of the heat and salt produced in freezing. Due to the high viscosity of the medium, turbulence is severely hindered. Mechanical diffusion induced by stirring of the ice by the wind and by gravity waves, on the contrary, may play an important role. A possible mechanism is shear-induced diffusion. It is possible to make some estimate in the two cases diffusion is generated by wind stirring and by gravity waves. The wind will generate in the ice layer a shear of strength \(a_s\approx (u_*)^2/\nu_{grease}\), where \(u_*\) is the friction velocity and \(\nu_{grease}\) is the effective viscosity of grease ice. In a concentrated suspension, such as grease ice, particle diffusion will arise from collisions, rather than from hydrodynamic interaction between particles. Collisions will occur between ice crystals travelling along flow lines separated by a typical crystal size \(L\), corresponding to a collision cross section \(\sigma\approx L^2\), and to a typical relative velocity in collision \(u_L\approx La_s\). In concentrated conditions, the number density of the crystals is \(n\approx L^{-3}\) and the particle displacement in a collision is \(\approx L\). Hence, the collision rate will be \(a_{coll}\approx n\sigma_{coll}u_L\approx a_s\), and the wind contribution to the diffusivity \(\kappa_{wind}\approx a_sL^2\). Taking a wind speed of about 10 m/s, corresponding to a friction velocity \(u_*\approx 0.01\) m/s, and taking for the grease ice, \(L\approx 1\) mm and \(\nu_{grease}\approx 0.01\ {\rm m^2/s}\), would give \[\begin{aligned}
\kappa^{wind}\approx (u_*L)^2/\nu_{grease}\approx 10^{-8}\ {\rm m^2/s}. \label{kappa_wind} \end{aligned}\] The contribution to diffusion from gravity waves can be estimated in a similar way. A crystal in the field of a gravity wave will be able to "prod" a nearby crystal only during the compressible phase of the wave at the crystal location. The resulting displacement will be \(S\approx u_L/\omega\approx La_s/\omega\), where \(\omega\) is the wave frequency, \(a_s\approx\epsilon\omega\) is the strain intensity, \(\epsilon=kU/\omega\) is the dimensionless wave amplitude, and \(k\) and \(U\) are the wavenumber and the characteristic fluid velocity of the wave. Hence \(S\approx\epsilon L\). In concentrated conditions, a crystal will make collision with a neighbour about once in a period, and if the degree of irreversibility in collision is sufficient, a contribution to the diffusivity, \(\kappa^{wave}\approx S^2\omega\), will be generated. Taking typical values for ocean waves, \(\omega\approx 1\ {\rm s^{-1}}\) and \(\epsilon\approx 0.1\), would give \[\begin{aligned}
\kappa^{wave}\approx (\epsilon L)^2\omega\approx 10^{-8}\ {\rm m^2/s}. \label{kappa_wave} \end{aligned}\] If molecular diffusion is small, tracers will move with the liquid around the particles, and have the same diffusion properties of the particles, described by Eqs. ([\[kappa_wind\]](#kappa_wind){reference-type="ref" reference="kappa_wind"}) and ([\[kappa_wave\]](#kappa_wave){reference-type="ref" reference="kappa_wave"}). The diffusivities in Eqs. ([\[kappa_wind\]](#kappa_wind){reference-type="ref" reference="kappa_wind"}) and ([\[kappa_wave\]](#kappa_wave){reference-type="ref" reference="kappa_wave"}) can then be compared with the heat and salinity diffusivities in salt water: \(\kappa_T\simeq 1.4\times 10^{-7}\ {\rm m^2/s}\) and \(\kappa_S\simeq 7.4\times 10^{-10}\ {\rm m^2/s}\). The mechanical contribution to diffusion appears to be relevant only for salinity. The estimates in Eqs. ([\[kappa_wind\]](#kappa_wind){reference-type="ref" reference="kappa_wind"}) and ([\[kappa_wave\]](#kappa_wave){reference-type="ref" reference="kappa_wave"}) disregard the possible impact of relative motions between phases, which may lead to diffusion phenomena analogous to those generated in flows in fixed particle beds. Relative motion of the phases is going to be generated in concentrated suspensions by the interplay between non-hydrodynamic and hydrodynamic (drag) forces on the particles. In creeping flow conditions, exact balance between mechanical and drag forces must exist, which corresponds, macroscopically, to balance between the drag forces by the interstitial flow and the resistance to deformation in the solid matrix. In diluted suspensions, no mechanical interparticle interactions obviously exist, and balance of solid and fluid stresses occurs at the particle scale, that is also the scale of the fluid velocity disturbances. In grease ice, non-hydrodynatic stresses are generated dynamically by the continuous formation and breaking of bonds between ice crystals induced by the flow, and allow the medium to persist in a fluid state. Note that in order for a crossflow to be present, a pressure gradient must exist in the liquid phase. Such pressure gradients are absent in the homogeneous shear layer generated by the wind stress. They are generated dynamically, instead, in the field of a gravity wave. The question remains whether the contribution to diffusion is significant or not. Answering this question is the goal of the present paper. To study the problem, a two-phase approach is adopted. No attempt is made to determine the macroscopic stresses of the network from microscopic behaviors, rather, an isotropic viscous response to deformation is assumed, with viscosity extrapolated from that of real grease ice. Knowing the relative velocity of the phases allows to obtain diffusivity estimates based on the characteristic size of the particles in suspension. Not more than an order of magnitude estimate is provided: grease ice is a complicated mixture of mm and sub-mm crystals in the form of platelets, needles and other shapes, with a solid volume fraction \(\bar C\approx 0.1\). Analytical treatment of the problem is therefore very difficult. More precise results on diffusion in the field of a gravity wave are obtained in the Appendix, in the idealized case of a dilute fiber network. The analysis predicts significant diffusivity enhancement with respect to the estimates in Eqs. ([\[kappa_wind\]](#kappa_wind){reference-type="ref" reference="kappa_wind"}) and ([\[kappa_wave\]](#kappa_wave){reference-type="ref" reference="kappa_wave"}), with an amplification factor proportional to the ratio between the distance traveled by a tracer in the solid matrix and the crystal scale. The key condition for amplification is that the effective viscosity of the slurry is large compared to that of the interstitial liquid. The diffusivity dependence on the solid volume fraction is weaker. The paper is organized as follows. In Sec. 2, the macroscopic equations describing the two-phase system are derived. In Sec. 3, the equations are applied to the calculation of the relative motion between phases. In Sec. 4, the contribution to diffusivity is evaluated. Section 5 is devoted to conclusions. Analytical results for diffusion in a dilute random fiber bed are presented in the Appendix.
# Two-phase model
The dynamics of gravity waves in grease ice covered ocean is studied by modelling the ocean as an infinitely deep uniform slurry. As the thickness of grease ice in the ocean, and the depth of the region where wave dissipation takes place, are usually comparable , the model is going to be rough, but not completely inaccurate. A macroscopic two-phase description is adopted, with the solid and liquid phases obeying separate mass and momentum conservation equations. No hypothesis is made for the moment on the microscopic structure of the medium, except for the fact that non-hydrodynamic interactions dominate the stress in the ice matrix. To fix the ideas, take the \(x\) axis along the direction of propagation of the wave, and the \(z\) axis pointing upwards, with \(z=0\) at the surface. The momentum and continuity equations obeyed by the two phases read: \[\begin{aligned}
&&(1-\bar C)[\rho_l\frac{\partial{\bf U}_l}{\partial t}+\nabla P-\mu\nabla^2{\bf U}_l]=\Gamma({\bf U}_s-{\bf U}_l)+(1-C)\rho_l{\bf g}, \label{U_l} \\ &&\bar C[\rho_s\frac{\partial{\bf U}_s}{\partial t}+\nabla P-\mu\nabla^2{\bf U}_l] =-\nabla P_s+\mu_s^{\scriptscriptstyle{B}}\nabla\nabla\cdot{\bf U}_s+\mu_s\nabla^2{\bf U}_s-\Gamma({\bf U}_s-{\bf U}_l)+C\rho_s{\bf g}, \label{U_s} \\ &&\frac{\partial\tilde C}{\partial t}+\bar C\nabla\cdot{\bf U}_s=0, \label{continuity} \\ &&\nabla\cdot[(1-\bar C){\bf U}_l+\bar C{\bf U}_s]=0, \label{incompressibility} \end{aligned}\] where linearity has been imposed, tilde and overbar indicate fluctuating and mean quantities, \(g\simeq 9.8\ {\rm m/s^2}\), \(\Gamma\) and \(C\) are the gravitational acceleration, the drag coefficient and the solid volume fraction, and \(\rho_n\), \({\bf U}_n\) and \(\mu_n\) are the mass density, the velocity and the dynamic viscosity of the liquid (\(n=l\)) and solid (\(n=s\)) phase. Similar equations have been used in to model the flow in deformable porous media. The terms \(\nabla P_s\), \(\mu_s^{\scriptscriptstyle{B}}\nabla\nabla\cdot{\bf U}_s\) and \(\mu_s\nabla^2{\bf U}_s\) in Eq. ([\[U_s\]](#U_s){reference-type="ref" reference="U_s"}) give the non-hydrodynamic part of the pressure and viscous forces. All three contributions arise from coarse graining of the particle mechanical interactions. Note that when these terms are zero, and the particles are neutrally buoyant, Eqs. ([\[U_l\]](#U_l){reference-type="ref" reference="U_l"}) and ([\[U_s\]](#U_s){reference-type="ref" reference="U_s"}) have solution \({\bf U}_l={\bf U}_s\). The mechanical pressure \(P_s\) and the bulk viscosity \(\mu_s^{\scriptscriptstyle{B}}\) account for the resistance to compression of the solid matrix. An isotropic expression is utilized for the bulk viscosity term even though such normal stresses are typically anisotropic in interacting particle suspensions. As the flow of the individual phases is almost incompressible (see below), the expected error is small. The terms \(\nabla P\) and \(\mu\nabla^2{\bf U}_l\) in Eqs. ([\[U_l\]](#U_l){reference-type="ref" reference="U_l"}) and ([\[U_s\]](#U_s){reference-type="ref" reference="U_s"}) give the hydrodynamic part of the pressure and viscous forces, with \(\mu\nabla^2{\bf U}_l\) receiving contribution from the stress both in the liquid and in the particles (the Einstein correction in a dilute suspension). The hydrodynamic pressure \(P\) is determined by the incompressibility condition Eq. ([\[incompressibility\]](#incompressibility){reference-type="ref" reference="incompressibility"}). The mechanical pressure is taken to obey a constitutive relation in the form \(P_s=P_s(C;\nabla{\bf U}_s)\), and to rapidly vanish for \(C>\bar C\), to allow hydrostatic equilibrium at \(C\approx\bar C\). Hydrostatic equilibrium for both phases is obtained by setting \({\bf U}_{l,s}=0\) in Eqs. ([\[U_l\]](#U_l){reference-type="ref" reference="U_l"}) and ([\[U_s\]](#U_s){reference-type="ref" reference="U_s"}). The result is, assuming slow variation of \(\bar C\) throughout the domain, \[\begin{aligned}
\bar P_s(z)=\bar P_s(0)+\bar Cg(\rho_l-\rho_s)z \quad{\rm and}\quad \bar P(z)=\bar P(0)-g\rho_lz. \label{buoyancy} \end{aligned}\] (Note that in the absence of a mechanical pressure, \(\bar P(z)=\bar P(0)-g[\rho_l+\bar C(\rho_s-\rho_l)]z\), buoyancy forces on the two phases, \(-\bar f_l=\bar f_s=-\bar C(\partial_z\bar P+\rho_sg)=g(1-\bar C)\bar C(\rho_l-\rho_s)\), would be present, and would need to be balanced by the drag forces generated by settling of the particles to the surface.) Subtraction of the hydrostatic component accounted for by Eq. ([\[buoyancy\]](#buoyancy){reference-type="ref" reference="buoyancy"}), from Eqs. ([\[U_l\]](#U_l){reference-type="ref" reference="U_l"}) and ([\[U_s\]](#U_s){reference-type="ref" reference="U_s"}), gives \[\begin{aligned}
&&(1-\bar C)[\rho_l\frac{\partial{\bf U}_l}{\partial t}+\nabla \tilde P-\mu\nabla^2{\bf U}_l]= \Gamma({\bf U}_s-{\bf U}_l), \label{U_l1} \\ &&\bar C[\rho_s\frac{\partial{\bf U}_s}{\partial t}+\nabla\tilde P-\mu\nabla^2{\bf U}_l] =-\beta\nabla\tilde C+\mu_s^{\scriptscriptstyle{B}}\nabla\nabla\cdot{\bf U}_s+\mu_s\nabla^2{\bf U}_s \nonumber \\ &&\qquad\qquad\qquad\qquad\qquad\qquad\quad\ -\Gamma({\bf U}_s-{\bf U}_l)+\tilde C(\rho_s-\rho_l){\bf g}, \label{U_s1} \end{aligned}\] where \[\begin{aligned}
\beta=(\partial P_s/\partial C)_{C=\bar C} \label{compressibility} \end{aligned}\] is the compressibility of the solid phase. In the case of monocromatic waves, one can write for the fluctuating quantities: \[\begin{aligned}
\tilde C({\bf r},t)=\tilde C(z){\rm e}^{{\rm i}(kx-\omega t)}, \qquad \tilde C(z)=\tilde c_p{\rm e}^{kz}+\sum_n\tilde c_n{\rm e}^{\alpha_nz},\ldots, \label{tilde C} \end{aligned}\] where subscripts \(p\) and \(n\) identify potential and non-potential components. The exponents \(\alpha_n\) give the depth of the boundary layers induced by the two-phase dynamics at the water surface. The velocities can be expressed in terms of velocity potentials: \[\begin{aligned}
{\bf U}_s=-\nabla\Phi-\nabla_\perp A, \qquad {\bf U}_l={\bf U}_s-\nabla\hat\Phi-\nabla_\perp\hat A, \label{potentials} \end{aligned}\] where \(\nabla_\perp=(\partial_z,-\partial_x)\). The fields \(A\) and \(\tilde A\) are chosen to contribute only to the vortical part of \({\bf U}_{l,s}\), and therefore do not have potential components \(A_p=a_p{\rm e}^{kz}\) and \(\hat A_p=\hat a_p{\rm e}^{kz}\). Such components could be eliminated anyway by a gauge transformation \(\Phi\to\Phi-{\rm i} A_p\), \(\hat\Phi\to\hat\Phi-{\rm i}\hat A_p\). Substitution of Eq. ([\[potentials\]](#potentials){reference-type="ref" reference="potentials"}) into Eqs. ([\[continuity\]](#continuity){reference-type="ref" reference="continuity"}) and ([\[incompressibility\]](#incompressibility){reference-type="ref" reference="incompressibility"}) gives \[\begin{aligned}
\tilde c_p= 0, \qquad \tilde c_n=-\frac{(\alpha_n^2-k^2)\bar C}{{\rm i}\omega}\phi_n \quad{\rm and}\quad \phi_n=-(1-\bar C)\hat\phi_n. \label{pippo0} \end{aligned}\] Equations for the non-potential components \(\tilde p_n\), \(\hat\phi_n\), \(a_n\) and \(\hat a_n\) are obtained by taking \(\nabla_\perp\) and \(\nabla\) of Eqs. ([\[U_l1\]](#U_l1){reference-type="ref" reference="U_l1"}) and ([\[U_s1\]](#U_s1){reference-type="ref" reference="U_s1"}), and then using Eq. ([\[pippo0\]](#pippo0){reference-type="ref" reference="pippo0"}): \[\begin{aligned}
&&(1-\bar C)[{\rm i}\omega\rho_l+\mu(\alpha_n^2-k^2)](a_n+\hat a_n)-\Gamma\hat a_n=0, \label{pippa1} \\ &&\bar C[{\rm i}\omega\rho_sa_n+\mu(a_n+\hat a_n)] +\mu_s(\alpha_n^2-k^2)a_n +\Gamma\hat a_n-\frac{g(\rho_l-\rho_s)\bar Ck}{\omega}(1-\bar C)\hat\phi_n=0, \label{pippa2} \\ && \{[{\rm i}\omega\rho_l+\mu(\alpha_n^2-k^2) ]\bar C(1-\bar C)-\Gamma\}\hat\phi_n+(1-\bar C)p_n=0, \label{pippa3} \\ &&\{[-{\rm i}\omega\rho_s\bar C-(\mu_s^{\scriptscriptstyle{B}}+\mu_s-\bar C\mu-\frac{\beta\bar C}{{\rm i}\omega})(\alpha_n^2-k^2) \nonumber \\ &&\qquad\quad-\frac{g(\rho_l-\rho_s)\bar C\alpha_n}{{\rm i}\omega}](1-\bar C)+\Gamma \}\hat\phi_n +\bar C\tilde p_n=0. \label{pippa4} \end{aligned}\] The potential terms are obtained by integrating Eqs. ([\[U_l\]](#U_l){reference-type="ref" reference="U_l"}) and ([\[U_s\]](#U_s){reference-type="ref" reference="U_s"}) over \(z\), and then exploiting Eqs. ([\[buoyancy\]](#buoyancy){reference-type="ref" reference="buoyancy"}) and ([\[pippo0\]](#pippo0){reference-type="ref" reference="pippo0"}-[\[pippa4\]](#pippa4){reference-type="ref" reference="pippa4"}). The result is \[\begin{aligned}
&& (1-\bar C)[{\rm i}\omega\rho_l(\phi_p+\hat\phi_p) +p_p+\rho_lg\eta_l]-\Gamma\hat\phi_p=0, \label{pippa5} \\ && \bar C[{\rm i}\omega\rho_s\phi_p+p_p+\rho_sg\eta_s]+\Gamma\hat\phi_p=0, \label{pippa6} \end{aligned}\] where the vertical displacement at the surface, \(\eta^{s,l}\), obeys \[\begin{aligned}
-{\rm i}\omega\eta^{s,l}=U_z^{s,l}|_{z=0}. \label{pippa7} \end{aligned}\] For weakly damped waves (which is the case in most situations of interest ), \(k\simeq k_\omega=\omega^2/g\), and the dominant contribution to \({\bf U}_l\simeq{\bf U}_s\) is produced by \(\phi_p\). The velocity scale of the waves is therefore \(U=k_\omega|\phi_p|\).
# Determining the crossflow
It is convenient to shift to dimensionless variables. The various quantities appearing in Eqs. ([\[U_l\]](#U_l){reference-type="ref" reference="U_l"}-[\[incompressibility\]](#incompressibility){reference-type="ref" reference="incompressibility"}) and below are rescaled as follows: \(\omega t\to t\), \(k/k_\omega\to k\), \(\mu/\mu_s\to\mu\), \(\sqrt{\mu_s/(\omega\rho_l)}\alpha\to\alpha\), \(\beta\bar C/(\omega\mu_s)\to\beta\), \(\Gamma/(\omega\rho_l)\to \Gamma\), \(pk_\omega/(\rho_lg)\to p\), \(Uk_\omega/\omega\to U\). The velocity scale of the waves becomes \(U=\epsilon\). The following dimensionless parameters are introduced: \[\begin{aligned}
\gamma=1+\frac{\mu_s^{\scriptscriptstyle{B}}}{\mu_s}, \qquad \rho=\frac{\rho_s}{\rho_l}, \quad{\rm and}\quad \nu_g= \frac{k_\omega^{3/2}\mu_s}{\rho_lg^{1/2}}, \label{dimensionless} \end{aligned}\] and the following smallness conditions are imposed \[\begin{aligned}
\mu,\ \nu_g,\ \bar C \ll 1. \label{small} \end{aligned}\] The stresses in the solid matrix are assumed to be of comparable magnitude; therefore \(\gamma,\beta=O(1)\). The conditions in Eq. ([\[small\]](#small){reference-type="ref" reference="small"}) correspond, in the order, to the requirements:
- high viscosity of the medium compared to that of the liquid phase;
- waves that are only weakly damped, in such a way that \(|k-1|\ll 1\) and all velocity potentials are small except \(\phi_p\);
- diluteness of the medium.
The first two conditions are easily satisfied by grease ice in the ocean. Assuming that all of the viscosity renormalization comes from mechanical interactions, in such a way that \(\mu\approx\mu_l= \rho_l\nu_l\), \(\nu_l\simeq 1.5\times 10^{-6}\ {\rm m^2/s}\), and taking for the viscosity in the solid matrix \(\nu_g\approx\nu_{grease}\approx 0.01\ {\rm m^2/s}\), would give \(\mu\approx 10^{-4}\). In similar fashion, considering wavelengths in the range of \(10-100\) m, would give for the dimensionless viscosity, \(1.5\times 10^{-3}>\nu_g>5\times 10^{-5}\). The last condition \(\bar C\ll 1\) is only marginally satisfied, and is adopted for simplicity. An upper bound for the grease ice concentration is fixed by the transition from grease ice to compact ice, which takes place at \(\bar C\approx 0.3\). Working with dimensionless variables and expanding to lowest order in the small quantities, Eqs. ([\[pippa1\]](#pippa1){reference-type="ref" reference="pippa1"}-[\[pippa4\]](#pippa4){reference-type="ref" reference="pippa4"}) take the form \[\begin{aligned}
&&{\rm i} a_n-\Gamma\hat a_n\simeq 0, \label{peppe1} \\ &&[{\rm i}\rho\bar C+\alpha_n^2]a_n +\Gamma\hat a_n-(1-\rho)\bar C\hat\phi_n\simeq 0, \label{peppe2} \\ &&\Gamma\hat\phi_n-p_n\simeq 0, \label{peppe3} \\ &&[{\rm i}(1-\rho)\nu_g^{-1/2}\bar C\alpha_n-(\gamma+{\rm i}\beta)\alpha^2_n +\Gamma]\hat\phi_n+\bar Cp_n\simeq 0 \label{peppe4} \end{aligned}\] From Eqs. ([\[peppe1\]](#peppe1){reference-type="ref" reference="peppe1"}-[\[peppe4\]](#peppe4){reference-type="ref" reference="peppe4"}) an incompressible mode for the individual phases is identified: [^1] \[\begin{aligned}
\alpha_i\simeq\sqrt{-{\rm i}}, \quad \hat a_i\simeq{\rm i} a_i/\Gamma, \quad \hat\phi_i\simeq p_i\simeq 0. \label{incompressible} \end{aligned}\] It is easy to recognize in this mode the standard behavior of the vortical component of the velocity field of a wave propagating in a viscous medium of viscosity \(\nu_g\). Equations ([\[peppe1\]](#peppe1){reference-type="ref" reference="peppe1"}-[\[peppe4\]](#peppe4){reference-type="ref" reference="peppe4"}) have an additional compressible mode with eigenmodes \[\begin{aligned}
\alpha_c\simeq\frac{{\rm i}(1-\rho)\bar C\pm\sqrt{4(\gamma+{\rm i}\beta)\Gamma\nu_g-(1-\rho)^2\bar C^2}} {2(\gamma+{\rm i}\beta)\sqrt{\nu_g}}, \label{alpha_c} \end{aligned}\] where only the root with a positive real part is admissible in an infinitely deep domain. The root's inverse gives the thickness of the compressible boundary layer. From Eqs. ([\[incompressible\]](#incompressible){reference-type="ref" reference="incompressible"}) and ([\[alpha_c\]](#alpha_c){reference-type="ref" reference="alpha_c"}), \(|\alpha_c/\alpha_i|\sim\max(\Gamma^{1/2},\bar C\nu_g^{-1/2})\gg 1\), which implies that the compressible boundary layer is exceedingly thin. From Eqs. ([\[peppe1\]](#peppe1){reference-type="ref" reference="peppe1"}-[\[peppe4\]](#peppe4){reference-type="ref" reference="peppe4"}) one can then write for the compressible component of the velocity potentials and of the pressure: \[\begin{aligned}
p_c\simeq \Gamma\hat\phi_c, \quad a_c\simeq \frac{(1-\rho)\bar C}{\alpha_c^2}\hat\phi_c, \quad \hat a_c\simeq \frac{{\rm i}(1-\rho)\bar C}{\alpha_c^2\Gamma}\hat\phi_c. \label{compressible} \end{aligned}\] To derive a dispersion relation, it is necessary to impose boundary conditions at the fluid free surface, and exploit Eqs. ([\[pippa5\]](#pippa5){reference-type="ref" reference="pippa5"}-[\[pippa7\]](#pippa7){reference-type="ref" reference="pippa7"}) to express the amplitude of non-potential modes in terms of potential ones. There are in total five unknowns: \(k\), \(p_p\), \(\hat\phi_p\), \(\hat\phi_c\) and \(a_i\); \(\phi_p\) remains arbitrary because of linearity. Thus, three boundary conditions are required. Two are well known from the study of gravity waves in viscous media, and are the requirement that the normal stress \(\tau_{zz}\simeq-\tilde P-\tilde P_s+2\mu_s\partial_zU_{sz}+\mu_s^{\scriptscriptstyle{B}}\nabla\cdot{\bf U}_s\) and the tangential stress \(\tau_{xz}\simeq\mu_s(\partial_xU_{sz}+\partial_zU_{sx})\) at the water surface are identically zero. A reasonable third boundary condition is that the interfaces with the atmosphere of the two phases move together, \([U_s-U_l]_{z=0}=0\), so that there are no sloshing phenomena at the water surface. Working with velocity potentials, \[\begin{aligned}
_{z=0}\simeq\hat\phi_p+ \alpha_c\nu_g^{-1/2}\hat\phi_c=0, \label{U-U} \end{aligned}\] and therefore, \[\begin{aligned}
\eta_s\simeq\eta_l\simeq-{\rm i} k\phi_p-a_i+{\rm i}\alpha_c\nu_g^{-1/2}\hat\phi_c. \label{eta} \end{aligned}\] Substitution of Eqs. ([\[U-U\]](#U-U){reference-type="ref" reference="U-U"}) and ([\[eta\]](#eta){reference-type="ref" reference="eta"}) into the equations for the potential terms, Eqs. ([\[pippa5\]](#pippa5){reference-type="ref" reference="pippa5"}) and ([\[pippa6\]](#pippa6){reference-type="ref" reference="pippa6"}), gives \[\begin{aligned}
&& {\rm i}(1-k)\phi_p+\tilde p_p-a_i+\Gamma\nu_g^{-1/2}\alpha_c\hat\phi_c=0, \\ &&\bar C[{\rm i}\rho(1-k)\phi_p+\tilde p_p-\rho a_i]-\Gamma\nu_g^{-1/2}\alpha_c\hat\phi_c=0, \end{aligned}\] which allows to determine the amplitude of the compressible mode as a function of potential modes: \[\begin{aligned}
\hat\phi_c\simeq\frac{(1-\rho)\bar C\nu_g^{1/2}}{\Gamma\alpha_c}\tilde p_p, \qquad \tilde p_p\simeq-{\rm i}(1-k)\phi_p+a_i. \label{hat phi_c} \end{aligned}\] The compressible mode is driven by buoyancy. The system of equations is closed by imposing the conditions on the normal and tangential stress at the surface: \[\begin{aligned}
&&\tau_{zz}|_{z=0}\simeq-\tilde p_p+2{\rm i}\nu_g^{1/2}\alpha_ia_i+[({\rm i}\beta+\gamma+1)\alpha_c^2+2{\rm i}(1-\rho)\nu_g^{1/2} \bar C\alpha_c^{-1}]\hat\phi_c=0, \label{tau_zz} \\ &&\tau_{xz}|_{z=0}\simeq-2{\rm i}\nu_g\phi_p+{\rm i} a_i+[(1-\rho)\bar C+2{\rm i} \nu_g^{1/2}\alpha_c]\hat\phi_c=0. \label{tau_xz} \end{aligned}\] Substitution of Eq. ([\[hat phi_c\]](#hat phi_c){reference-type="ref" reference="hat phi_c"}) into Eqs. ([\[tau_zz\]](#tau_zz){reference-type="ref" reference="tau_zz"}) and ([\[tau_xz\]](#tau_xz){reference-type="ref" reference="tau_xz"}) gives finally \[\begin{aligned}
k\simeq 1+4{\rm i}\nu_g, \qquad a_i\simeq 2\nu_g\phi_p, \qquad \hat\phi_c\simeq-\frac{2(1-\rho)\bar C\nu_g^{3/2}}{\Gamma\alpha_c}\phi_p. \label{reldisp} \end{aligned}\] One recovers the dispersion relation and the equation for the boundary layer structure of gravity waves in a homogeneous medium of kinematic viscosity \(\nu_g\simeq\mu_s/\rho_l\). The suspension inherits the viscosity of the solid phase, as expected. Similar mechanisms have been invoked in to explain the rheology of concentrated sphere suspensions. The compressible mode does not contribute to lowest order to the dynamics, as can be verified by back substitution of Eqs. ([\[reldisp\]](#reldisp){reference-type="ref" reference="reldisp"}) and ([\[alpha_c\]](#alpha_c){reference-type="ref" reference="alpha_c"}) into Eqs. ([\[tau_zz\]](#tau_zz){reference-type="ref" reference="tau_zz"}) and ([\[tau_xz\]](#tau_xz){reference-type="ref" reference="tau_xz"}). This suggests that use of a boundary condition on \(\eta^{s,l}\) could be avoided by simply setting \(\hat\phi_c=0\), thus neglecting all contributions from the compressible mode. The thin compressible boundary layer at the surface would be replaced in this way by a finite gap between phases, whose amplitude can be evaluated from Eqs. ([\[U-U\]](#U-U){reference-type="ref" reference="U-U"}) and ([\[reldisp\]](#reldisp){reference-type="ref" reference="reldisp"}): \(\eta_s-\eta_l\simeq 2(1-\rho)\bar C\nu_g\Gamma^{-1}\phi_p\). It is now possible to determine the relative motion of the phases. Inspection of Eqs. ([\[incompressible\]](#incompressible){reference-type="ref" reference="incompressible"}), ([\[compressible\]](#compressible){reference-type="ref" reference="compressible"}), ([\[U-U\]](#U-U){reference-type="ref" reference="U-U"}) and ([\[reldisp\]](#reldisp){reference-type="ref" reference="reldisp"}) allows to conclude that the dominant contribution is produced by the vortical incompressible mode \(\hat a_i\). This leads to the final result \[\begin{aligned}
{\bf U}_l-{\bf U}_s\simeq-({\rm i}\sqrt{-{\rm i}},\nu_g^{1/2})\frac{2\nu_g^{1/2}}{\Gamma}\phi_p. \label{U_l-U_s} \end{aligned}\] The relative motion between phases is predominantly along the horizontal, and is confined in a region of thickness \(\lambda_\alpha\approx\nu_g^{1/2}\) near the surface. The incompressible mode dominates the dynamics both with regard to the corrections to wave propagation (through \(a_i\)) and to the interaction between solid and liquid phase (through \(\hat a_i\)).
# Diffusivity estimate
Knowing the crossflow \({\bf U}_l-{\bf U}_s\) allows to estimate the medium diffusivity in terms of the characteristic scale \(L\) of the crystals \[\begin{aligned}
\kappa^{wave}\approx |{\bf U}_l-{\bf U}_s|L. \label{kappa} \end{aligned}\] A difficulty arises however because of the oscillatory character of \({\bf U}_l-{\bf U}_s\). In creeping flow conditions, reversibility of Stokes flows would lead, in the absence of mechanical interactions of the crystals, to pure oscillatory motion of the crystals and of the interstitial liquid; tracer trajectories would be closed and no diffusion would be possible. In order to have diffusion, the effect of inertia and of mechanical interactions must be taken into account, with chaos in the tracer trajectories possibly helping in the process. In this regards, arguments in suggest that tracer excursions in the solid matrix of the order of just a few crystal spacings could be enough to produce trajectory decorrelation. If a condition of large tracer excursion compared to the crystal spacing is satisfied, \(\lambda\approx\omega^{-1}|{\bf U}_l-{\bf U}_s|\gg L\), diffusive behaviours are therefore expected, and Eq. ([\[kappa\]](#kappa){reference-type="ref" reference="kappa"}) applies. Considering waves of normalized amplitude \(\epsilon\), substitution of Eq. ([\[U_l-U_s\]](#U_l-U_s){reference-type="ref" reference="U_l-U_s"}) into Eq. ([\[kappa\]](#kappa){reference-type="ref" reference="kappa"}) gives, after making \(L\) dimensionless by rescaling \(k_\omega L\to L\), \[\begin{aligned}
\kappa^{wave}\approx \frac{\nu_g^{1/2}L\epsilon}{\Gamma}. \label{kappa_zz} \end{aligned}\] The large excursion condition \(\lambda\gg L\) becomes, using again Eq. ([\[U_l-U_s\]](#U_l-U_s){reference-type="ref" reference="U_l-U_s"}), \[\begin{aligned}
\nu_g^{1/2}\epsilon\gg\Gamma. \label{cond0} \end{aligned}\] The drag coefficient in Eqs. ([\[kappa_zz\]](#kappa_zz){reference-type="ref" reference="kappa_zz"}) and ([\[cond0\]](#cond0){reference-type="ref" reference="cond0"}) can be estimated as the product \(\Gamma\approx nF\) of the drag force on an individual particle \(F\approx \mu_lL\) and the particle density \(n\approx L^{-3}\): \[\begin{aligned}
\Gamma\approx \frac{\mu\nu_g}{L^2}. \label{Gamma} \end{aligned}\] The condition \(\lambda\gg L\), Eq. ([\[cond0\]](#cond0){reference-type="ref" reference="cond0"}), becomes therefore \[\begin{aligned}
{\cal R}=\frac{\lambda}{L}\approx\frac{\epsilon L}{\mu\nu_g^{1/2}}\gg 1, \label{calR} \end{aligned}\] and the diffusivity in Eq. ([\[kappa_zz\]](#kappa_zz){reference-type="ref" reference="kappa_zz"}), back to dimensional units, takes the form \[\begin{aligned}
\kappa^{wave}\approx {\cal R} L^2\omega. \qquad \label{kappa_zz1} \end{aligned}\] The amplification factor with respect to the shear-induced part of the diffusion coefficient in Eq. ([\[kappa_wave\]](#kappa_wave){reference-type="ref" reference="kappa_wave"}) is \({\cal R}/\epsilon^2\). The diffusivity is independent of the volume fraction \(\bar C\). A more precise calculation carried out in the Appendix, in the case of a dilute fiber suspension, confirms the result within logarithmic corrections. The estimate in Eq. ([\[kappa_zz1\]](#kappa_zz1){reference-type="ref" reference="kappa_zz1"}) can be applied to the case of gravity waves in grease ice. In the case of 100 m long waves, 1 mm ice crystals, and an ice viscosity \(\nu_g\approx\nu_{grease}\approx 0.01\ {\rm m^2/s}\), corresponding in dimensionless variables to \(\nu_g\approx 5\times 10^{-5}\) and \(\Gamma\approx 2\), Eqs. ([\[calR\]](#calR){reference-type="ref" reference="calR"}) and ([\[kappa_zz1\]](#kappa_zz1){reference-type="ref" reference="kappa_zz1"}) would give \[\begin{aligned}
{\cal R}\approx 15, \quad{\rm and}\quad \kappa^{wave}\approx 10^{-5}\ {\rm m^2/s}. \label{calR1} \end{aligned}\] Using the same ice parameters in the case of 10 m waves, corresponding in dimensionless variables to \(\nu_g\approx 1.5\times 10^{-3}\) and \(\Gamma\approx 0.6\), would give instead \({\cal R}\approx 16\) and \(\kappa^{wave}\approx 4\times 10^{-5}\ {\rm m^2/s}\). Note that the diffusivity increment described in Eq. ([\[calR1\]](#calR1){reference-type="ref" reference="calR1"}) is localized in the viscous boundary layer of thickness \(\lambda_\alpha=\sqrt{\nu_g/\omega}\) near the water surface, which, as already remarked, coincides in most situations of interest with the thickness of the grease ice layer.
# Concluding remarks
The analysis carried out in the present paper predicts a mechanical contribution to heat and mass diffusion in grease ice stirred by a gravity wave, that exceeds by several orders of magnitude the contribution by shear-induced diffusion. The corresponding increment over the molecular diffusivity is \(O(10^5)\) in the case of salinity, \(O(10^2)\) in the case of heat. Analysis in the Appendix suggests that the contribution of prefactors not included in the analysis may modify the estimates by \(O(10^{-2})\) factors, thus it is not possible to conclude that heat diffusion is strongly affected. Less likely that the prefactor contribution is so huge to lead to similar conclusions in the case of salinity. The predicted enhancement of haline diffusion may lead to a shift in the onset of haline convection during ice formation. The mechanism of diffusion described in this paper is not confined to suspension stirring by gravity waves. The mechanism is expected to be active in all situations in which pressure gradients in the flow and non-hydrodynamic interactions of the particles are important. A question which arises in the case of periodic flows is how much does Stokes flow reversibility in creeping flow conditions affect the development of diffusion. A similar question has attracted recently great attention in the case of shear-induced particle diffusion in periodic shear flows. In that case, irreversibility is restored by particle collisions, provided the shear is above a concentration dependent threshold. In the present case, it is reasonable to expect that a combination of chaos in the tracer trajectories, particle inertia, and dislocations induced by the flow, all play a role in producing the necessary irreversibility for tracer diffusion. The observation is strengthened by the predicted dependence of the diffusivity on the tracer excursion in a wave period in the solid bed compared to the crystal separation. Whether or not some threshold for irreversibility exists, analogous to the one in the case of shear-induced diffusion, remains to be ascertained. : This research was supported by FP7 EU project ICE-ARC (Grant agreement No. 603887).
[^1]: Note that in rescaled variables a factor \(\nu^{-1/2}_g\) is extracted out of \(\alpha_n\), in such a way that the \(z\)-derivative of non-potential fields \(Q_n\) is \(\partial_zQ_n=\nu_g^{-1/2}\alpha_nQ_n\). | {'timestamp': '2018-06-08T02:10:58', 'yymm': '1706', 'arxiv_id': '1706.01697', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01697'} |
null | null |
# Introduction {#sec:intro}
The one-point probability distribution function (PDF) of the matter density field in the universe, and its related statistics the distribution of galaxy counts, have a long and somewhat patchy history in cosmology and extragalactic astronomy. It was Edwin Hubble almost a century ago who found that the counts of about \(44,000\) extra-galactic nebulae distributed over a large area of the sky have a probability distribution that is not Gaussian but can be approximated by a log-normal distribution. The statistics of galaxy counts in the Lick survey, in projected cells of size \(10'\times 10'\), was studied by who also discovered that the distribution of the counts is much broader than the Poisson PDF. The \(rms\) of galaxy counts \(\sigma\) in cells of size \(R\) is an integral over the power spectrum of the galaxy distribution. As such, in former times, a count-in-cells analysis of the IRAS redshift galaxy survey was performed by who used the counts as a measure of the two-point clustering statistics on different scales. Once methods to estimate the correlation function and the power spectrum were developed and new large-scale galaxy surveys were available, the count-in-cells as clustering statistics started to play a secondary role. Higher moments of cell counts depend on correlation functions of order larger than two. This means that the whole PDF has information not only on the two-point clustering but also on higher order statistics, which by itself is very valuable information. At present, a precise description and modeling of the underlying matter density distribution-and biasing prescription that connects the dark matter field with the galaxy distribution-are fundamental to extract cosmological information from current and upcoming large-scale redshift and lensing galaxy surveys. For this reason in the last years there has been a rejuvenated interest in the cosmic density distribution both from cosmological \(N\)-body simulations and galaxy surveys. estimated the PDF of galaxies in the 2dF redshift survey using about 200,000 galaxies. Because of a relatively small volume, their analysis was done only for large cells of size 10-30 Mpc. They found that the log-normal distribution fits the data reasonably well, but the noise in the data did not allow them to make accurate measurements of the PDF. The situation was improved by using the count-in-cells statistics for galaxies in the SDSS main sample. They used \(\sim 100,000\) galaxies and estimated the PDF for spheres of radius \(R=(8-24)\mbox{\)h^{-1}\(Mpc}\). They found that the log-normal distribution was very inaccurate (a factor of \(\sim 2\) errors) for spheres of \(R=8\mbox{\)h^{-1}\(Mpc}\). A modification of the log-normal distribution (called log-normal+bias) somewhat improved the fits, but it still had \(\sim 50\%\) errors at low number of galaxy counts. The negative binomial distribution was a much better fit for all filtering scales. At hight redshift, studied the count-in-cells distribution of \(\sim 30,000\) galaxies in the VIPERS redshift survey with the typical average number of galaxies per cell of 0.5-5 and spherical cells of radius \(R=(4-8)\mbox{\)h^{-1}\(Mpc}\). They found that the skewed log-normal distribution (a modification of the log-normal distribution with 4 more free parameters) was not accurate enough to fit the results of observations. Instead, they found that the negative binomial distribution was much more accurate. Yet, using the DES science verification data confirmed that the log-normal model is a good fit to both the galaxy density contrast and weak lensing convergence PDFs on scales of \((3-10)\) Mpc at median redshift \(z = 0.3\). In spite of the fact that at present these seems to be the best observational results, the errors and noise in the data are still substantial. On the theoretical side the situation is also complicated. There are two types of approaches: models that start with some dynamical description of the non-linear evolution of the density field and proceed to make predictions of the matter PDF [e.g., @Juan1991; @Bernardeau1994; @Kofman1994; @Juan2002; @Ohta2003; @LamSheth2008]; and then there are phenomenological approximations that assume a specific analytical form of the PDF, and find best-fitted parameters of this distribution function to simulation data. Theoretical models based on the nonlinear dynamics typically use either some variant of the spherical infall model or the Zeldovich approximation. These models have made substantial progress and now can make very accurate predictions for relatively large smoothing scales \(R\gtrsim 5\mbox{\)h^{-1}\(Mpc}\) and small \(\sigma\lesssim 1\) giving errors less than \(\sim 10\%\). Not surprisingly, as expected, the models are not very useful and start to fail at larger \(rms\) fluctuations \(\sigma\gtrsim 1\). One of the disadvantages of the dynamical models is their complexity. They typically require some manipulation of the linear power spectrum, analytical approximations for different terms, and can be quite cumbersome to deal with. This is not a serious impediment to their use, but it is a nuisance. Simple analytical functions can serve as an alternative to more complicated dynamical models. The log-normal distribution is an example of this approach. It was heavily advocated by and is often used for relatively large smoothing scales. There is little justification why the density distribution function should be log-normal. argue that under the assumption that the divergence of the peculiar velocity field in Eulerian coordinates grows as the velocities themselves (as given by linear theory) the density field can be expressed as the exponential of a Gaussian field. But while their assumption is acceptable for the Lagrangian divergence, for the Eulerian one there is an additional growth roughly proportional to the cubic root of the normalized density. This leads to a density field that is equal to a Gaussian field to the third power, whose PDF is quite different from a log-normal distribution. Here is the main argument of: "The lognormal is one of the simplest ways of defining a fully self-consistent random field which always has \(\rho>0\) and, most importantly, is one of the few non-Gaussian random fields for which interesting properties are calculable analytically." This says that the PDF should be log-normal because it can be handled analytically--hardly a serious argument. Another argument is of the same caliber: the log-normal distribution is well-known and frequently used in other fields of science. Similar arguments were used for other phenomenological models. The only real justification for the existing phenomenological approximations (including the log-normal) is that they make a fit to \(N\)-body results. This is the reason why cosmological \(N\)-body simulations are important for the field. In this paper we use a very large suite of cosmological simulations to produce accurate estimates of the dark matter distribution functions. Our simulations cover a wide range of numerical and cosmological parameters. We use the estimates to test different dynamical models and approximations for the PDF and to study its dependance on redshift and cosmological parameters. One of our goals in this paper is to make a comprehensive study of the different effects that can be associated with \(N\)-body results regarding the matter density distribution, such as mass and force resolution, size of the box, shot noise and cosmic variance. In this regard we find that systematic errors in the PDF can be important. For example, noise related to the discreteness of the density probed by particles is the leading factor of seriously wrong estimates of the PDF in underdense regions. Generation of mock galaxy catalogs provides a motivation for our study of the density distribution function. One needs to produce thousands of realizations of the dark matter density and velocity fields. This can be done by carefully tuning parameters of simulations and limiting their resolution to a fraction of a megaparsec. A biasing prescription then connects the dark matter with galaxies. This path requires knowledge of the distribution of dark matter mass on very small scales \(100\mbox{\)h^{-1}\(kpc}-1\mbox{\)h^{-1}\(Mpc}\). This is a challenge because the resolution of these simulations is not sufficient to resolve individual halos and subhalos making it difficult to apply existing tools such as Halo Abundance Matching and Halo Occupation Distribution. A path to solve the problem is to map dark matter to galaxies using a biasing scheme which requires the understanding details of the density distribution function and finding limitations to its estimates. Unfortunately, only very few studies in the literature provide PDF results for small smoothing scales \(\lesssim 1\,\mbox{\)h^{-1}\(Mpc}\). So, we will make an effort to study this regime too. finds that in the regime of small smoothing and large density the PDF has a power-law shape with a slope of \(\approx-1.9\), which is similar to what we find in this work. This paper is organized as follows. In Sec. [2](#sec:def){reference-type="ref" reference="sec:def"} we define quantities related with the PDF and provide details of some analytical approximations in Sec. [3.1](#sec:LogN){reference-type="ref" reference="sec:LogN"}. The spherical infall and the double-exponential models are introduced in Sec. [3.2](#sec:sph){reference-type="ref" reference="sec:sph"}. Numerical simulations used in this paper are discussed in Sec. [4](#sec:sim){reference-type="ref" reference="sec:sim"}. Sec. [5](#sec:main){reference-type="ref" reference="sec:main"} also presents main features of the PDF. Accuracy of different approximations are discussed in Sec. [6](#sec:sphLog){reference-type="ref" reference="sec:sphLog"}. Summary of results is given in Sec. [7](#sec:concl){reference-type="ref" reference="sec:concl"}. Finally, numerical effects are discussed in Appendix [\[sec:cosmic\]](#sec:cosmic){reference-type="ref" reference="sec:cosmic"}, and Appendix B presents tables of parameters for the double-exponential approximation.
# Definitions {#sec:def}
In order to estimate the density distribution function \(P(\rho)\) from \(N\)-body simulations we split the computational volume \(L^3\), where \(L\) is the box size, with a \(3D\) mesh of size \(N^3_{\rm cell}\) and use the Cloud-In-Cell (CIC) density assignment scheme to estimate the density \(\rho\) at each grid point of the mesh. The cell size of the grid \(\Delta x =L/N_{\rm cell}\) defines the smoothing length. The density is normalized to the average matter density \(\rho_{av} \equiv \Omega_m\rho_{\rm cr}\), i.e. \[\rho \equiv 1+\delta_{\rm NL} = \rho_{\rm DM}/\Omega_m\rho_{\rm cr}, \label{eq:overdensity}\] where \(\delta_{\rm NL}\) is the matter density contrast or overdensity. The index \(NL\) highlights the fact that \(\rho\) is a non-linear quantity and it can be distinguished from the density contrast \(\delta\) as estimated by the linear theory. Throughout the paper we use the quantity \(\rho\) as "density" in spite of the fact that it is really a normalized density--a dimensionless quantity as shown in eq. ([\[eq:overdensity\]](#eq:overdensity){reference-type="ref" reference="eq:overdensity"}). This is done for convenience to avoid repeating \(1+\delta_{\rm NL}\) in most plots and equations. The values of density \(\rho\) are binned using logarithmically spaced bins with width \(\Delta\log_{10}(\rho) =0.025-0.050\). The density distribution function-PDF of the cosmic density field-is then defined as a normalized number of cells with density in the range \([\rho,\rho+\Delta\rho]\): \[P(\rho) = \frac{\Delta N_{\rm cell}}{N^3_{cell}\Delta\rho}. \label{eq:PDF}\] The PDF can have a surprisingly large range of values. For example, density can reach values larger than \(10^5\) for hundreds of cells when we use a large mesh of \(\sim 3000^3\) cells in high-resolution simulations. That gives \(P(\rho) \sim 10^{-12}\). At the same time the number of cells at low densities can be millions for a small density bin leading to a large PDF value \(P(\rho)\gtrsim 1\). In order to avoid a large dynamical range of quantities, we typically plot \(\rho^2P(\rho)\). By design, the density distribution function is normalized to have the total volume and total mass equal to unity: \[\int^\infty_0 P(\rho)d\rho =1, \quad \int^\infty_0 \rho P(\rho)d\rho =1. \label{eq:normalization}\] The second moment of \(P(\rho)\) gives the \(rms\) density fluctuation of the density field \(\sigma\), and as such it is related to the non-linear power spectrum \(P_{NL}(k)\) of density perturbations: \[\sigma^2 = \int^\infty_0 (\rho-1)^2 P(\rho)d\rho = \frac{1}{2\pi^2}\int^{k_{Ny}}_0P_{NL}(k)W^2(k\Delta x)k^2dk, \label{eq:sigma}\] where \(W^2(k\Delta x)\) is the power spectrum of the CIC filter with the width \(\Delta x\), and the integral is truncated at the Nyquist frequency of the mesh \(k_{Ny}=\pi/\Delta x\). Figure [\[fig:pdf\]](#fig:pdf){reference-type="ref" reference="fig:pdf"} shows our first results on the structure of the PDF. More detailed discussion is given later in Sec. [5](#sec:main){reference-type="ref" reference="sec:main"}. The plot demonstrates the main trend of the shape of the PDF. In the linear regime of the growth of fluctuations, when \(\sigma\ll 1\), the PDF is a Gaussian distribution that quickly acquires a skewed shape as \(\sigma\) increases. With the further increase of \(\sigma\) the PDF becomes progressively wider and becomes a power-law with the slope close to-2 that is smoothly suppressed on small and large densities.
# Methodology {#sec:method}
Different analytical approximations and theoretical models, as mentioned in Sec. [1](#sec:intro){reference-type="ref" reference="sec:intro"}, are used to fit and make predictions for the PDF estimates obtained from numerical simulations. Here we describe both approaches.
## PDF analytical approximations: log-normal, Negative Binomial, and Generalized Extreme Value {#sec:LogN}
We introduce the log-normal, negative binomial and the generalized extreme value distributions which have been traditionally adopted as analytical approximations for the PDF in many works. The **log-normal** (LN) distribution function \(P_{LN}\) is defined as: \[\rho P_{LN}(\rho) = \frac{1}{\sqrt{2\pi\sigma^2_{LN}}} \exp\left(-\frac{[\ln(\rho)+\sigma^2_{LN}/2]^2}{2\sigma^2_{LN}} \right), \label{eq:LN}\] where \[\sigma^2_{LN} =\ln[1+\sigma^2].\] is the only free parameter. \(\sigma^2_{LN}\) can be obtained from results of simulations, and thus should be considered as fixed. Because the log-normal distribution does not provide accurate fits to numerical simulations, a number of modifications have been proposed. None of those modifications extend the approximation to large densities, hence we do not discuss them in this paper. The **negative binomial** (NBN) distribution is defined as a discrete distribution. It is the probability \(P_N(V)\) to find \(N\) particles in a cell of volume \(V\) with the average number of particles \(\bar N\). It can be re-written as a distribution function of density contrast \(\rho =N/\bar N\): \[P_{NBN}(\rho) = \frac{1}{\bar N}\cdot \frac{\Gamma(N+1/g)}{\Gamma(1/g)\Gamma(N+1)} \cdot\frac{(g\bar N)^N}{(1+g\bar N)^{N+1/g}}, \label{eq:nbn}\] where \(g\) is a parameter that is defined by \(rms\) fluctuations of counts: \(g=(\sigma_N^2-\bar N)/\bar N^2\). The average number of particles per cell \(\bar N\) in general is not an integer number and is defined by the average density and cell volume: \(\bar N = \bar\rho V\). So, the two parameters \(g\) and \(\bar N\) that define the NBN distribution are not free and can be fixed from simulations. However, for PDFs with not too small \(rms\) fluctuations (\(\sigma\gtrsim 1\)) the fits produced by this distribution are not very accurate. As the result, we decided to treat both \(\bar N\) and \(g\) as free parameters. While the negative bimodal distribution formally has two parameters, there is little change in \(P_{NBN}\) when \(\bar N>10\). For cases where NBN makes some reasonable fits, our simulations have typical values of \(\bar N\) of many hundreds. So, in practice the NBN PDF depends only on one parameter \(g\), which defines the width of the distribution function: the larger is \(g\), the wider is \(P(\rho)\). At large average number of objects in a cell \(\bar N\gg 1\) and for large \(N>\bar N\) the NBN approximation predicts that the density distribution changes with density \(\rho = N/\bar N\) as follows \[P_{NBN}(\rho) \approx \frac{1}{\bar N^{1/g}\Gamma(1/g)} \rho^{1-1/g}\exp\left(-\frac{\rho}{g}\right).\] This expression is very different compared to the behavior of the PDF observed in \(N\)-body simulations for large \(\rho\) and \(\sigma\): in that regime \(P(\rho)\propto \rho^{-2}\exp\left(-C\rho^{0.5}\right)\). Thus, the NBN approximation predicts too steep a decline with density and lacks the power-law regime of the \(N\)-body PDF. A **generalized extreme value** (GEV) distribution as an approximation for the density distribution function was used by. It can be written as: \[\rho P_{GEV}(\rho) = \frac{1}{\ln(10)\beta}\frac{\exp(-z^{-1/k})}{z^{1+1/k}}, \, z\equiv 1+\frac{k}{\beta}\lg_{10}\left(\frac{\rho}{\rho_0}\right), \label{eq:GEV}\] where \(k, \rho_0,\) and \(\beta\) are free parameters.
## Spherical Infall and double-exponential PDF models {#sec:sph}
Approximations discussed so far were not based on any dynamical models. They simply make a guess regarding the functional form of \(P(\rho)\) and then proceed to finding parameters that produce the best fit. The guess is not based on any insights from the dynamics of clustering either. **Spherical infall models** are different because they are theoretical predictions for the density distribution function that are based on simplified approximations of the non-linear evolution of the density field. Here we closely follow the theoretical framework developed by and. We assume that the linear density field was smoothed with a top-hat filter with radius \(R_f\) and corresponding mass \(M\). The variance of the smoothed field is equal to \[\sigma_L^2(M) = \frac{1}{2\pi^2}\int k^2dk P(k)W^2(kR). \label{eq:sigmaL}\] In the spherical infall model the mapping from linear density contrast \(\delta_L\) to the nonlinear overdensity \(\rho\) is approximated by the following relation: \[\rho = \left(1-\frac{\delta_L}{\delta_c}\right)^{-\delta_c}, \quad \rho\equiv \frac{M}{\bar M},\quad \bar M=\rho_b\Delta x^3,\] where \(\delta_c\) is the linear theory prediction for the critical overdensity of collapse. Here we will use \(\delta_c=5/3\) as suggested by. For the initial Gaussian fluctuations this model gives the density distribution function: \[\rho^2P(\rho) =\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\delta_L^2}{2\sigma_L^2}\right) \frac{d(\delta_L/\sigma_L)}{d\ln\rho}. \label{eq:sph1}\] This is the same as eq.(6) in. We also use a modification of the spherical infall model, which is based on the excursion set model: \[\rho^2P(\rho) =\frac{1}{\sqrt{2\pi\sigma_L^2}}\exp\left(-\frac{\delta_L^2}{2\sigma_L^2}\right) \left[1-\delta_L\left(\frac{1}{\delta_c}-\frac{\gamma}{3}\right) \right], \label{eq:sph2}\] where \(\gamma(\rho)\) is: \[\gamma =-3\frac{d\ln\sigma_L^2}{d\ln M}. \label{eq:gam}\] Note that in these equations \(\sigma_L\) is a function of filtering mass \(M\), which in turn depends on density \(\rho = M/\bar M\). So, \(\sigma_L = \sigma_L(\rho)\). In order to apply the models, we need to adjust the top-hat filtering scale \(R_f\) used in the spherical infall model so that it matches the Cloud-In-Cell filtering scale used in our simulations. This is done by matching the power spectra of both filters at wavenumbers \(k<0.7k_{Ny}\) by applying \(R_f=\Delta x/\sqrt{1.3}\), where \(k_{Ny}=\pi/\Delta x\) is the Nyquist frequency of the density grid used for density assignment, and \(\Delta x\) is the grid size. Specifically, for every bin with density \(\rho\) we find the mass \(M=\rho\bar M\) and then the top-hat filtering radius is found as \(R_f= \sqrt{2}\left[3M/4\pi\rho_{cr}\Omega_m\right]^{1/3}\). When applying the relations given in eqs.([\[eq:sigmaL\]](#eq:sigmaL){reference-type="ref" reference="eq:sigmaL"}--[\[eq:gam\]](#eq:gam){reference-type="ref" reference="eq:gam"}), we integrate numerically eq.([\[eq:sigma\]](#eq:sigma){reference-type="ref" reference="eq:sigma"}) with the top-hat filter \(W(kR_f)\) and use finite differences to estimate the derivatives in eq.([\[eq:sph1\]](#eq:sph1){reference-type="ref" reference="eq:sph1"}) and eq.([\[eq:gam\]](#eq:gam){reference-type="ref" reference="eq:gam"}).
:::
As was found previously [e.g., @Juan2002; @LamSheth2008; @LamShethEllips2008; @Neyrinck2016], the spherical infall model provides good approximations for the PDF in those cases with large filtering scales where the \(rms\) fluctuation in the simulation box \(\sigma(M)\) is relatively small \(\sigma\lesssim 1\). This is consistent with our results, which will be presented later. At larger \(\sigma\) (large densities) the model provides results that are much less accurate. Even so, the spherical infall model predicts trends that track our \(N\)-body results. This is somewhat unexpected because at large overdensities \(\rho \gtrsim 10^2\) the model is clearly outside of the limits of its dynamical applicability. After all, a simplistic treatment of non-linear evolution used by the model cannot be valid for densities that are appropriate for collapsed and virialized halos. It is interesting and instructive to find what makes the spherical infall model much better than expected. The last factor in eq. ([\[eq:sph1\]](#eq:sph1){reference-type="ref" reference="eq:sph1"}) is just a correction to the leading terms that are the power-law \(P\propto \rho^{-2}\) (see also Figure [\[fig:pdf\]](#fig:pdf){reference-type="ref" reference="fig:pdf"}) and the exponential term on the right-hand-side of eq. ([\[eq:sph1\]](#eq:sph1){reference-type="ref" reference="eq:sph1"}). The exponential term originates from the assumption that the density distribution function of primordial fluctuations is Gaussian. The \(\rho^{-2}\) term comes from integrating the Gaussian distribution function over mass and then by writing it in differential form. This is basically the same logic as in the Press-Schechter derivation of the mass function of dark matter halos. More in detail, we see that the exponential term provides the truncation of the \(P\propto \rho^{-2}\) behavior both on the low-density \(\rho<1\) and on the high-density \(\rho\gg 1\) regimes. At low densities (small masses, large \(\sigma_L\)) the truncation is mostly due to large negative values of \(\delta_L\). At large \(\rho\) (large masses and small \(\sigma_L\)) the decline is related to the combination of decreasing \(\sigma_L\) and increasing \(\delta_L\). However, the increase of \(\delta_L\) is limited: it can not exceed \(\delta_{L,{\rm max}}=\delta_c=5/3\), and the decline in \(\sigma_L\) is not strong enough by itself to produce a substantial suppression of the PDF. So, the \(\rho^{2}P(\rho)\) shape becomes nearly flat at very large densities \(\rho > 100\) and small smoothing scales \(\Delta x \lesssim 1\mbox{\)h^{-1}\(Mpc}\). This is clearly seen in Figure [\[fig:pdf2\]](#fig:pdf2){reference-type="ref" reference="fig:pdf2"}. At very small smoothing scales \(\Delta x \ll 1\mbox{\)h^{-1}\(Mpc}\) and large densities there is another regime, that sheds light on how the density distribution function should behave at very large densities. When density is larger than \(\sim 100\) we are likely dealing with interiors of collapsed dark matter halos. In this regime the distribution function \(P(\rho)\) is a sum over the distribution functions of individual halos. Assuming that the density profile of a halo can be approximated by the Navarro-Frank-White (NFW) profile, and for a very small filtering scale, we can can derive the PDF provided by a single halo. If \(\rho(r)\) is the halo density profile, then the density distribution function given in eq. ([\[eq:PDF\]](#eq:PDF){reference-type="ref" reference="eq:PDF"}) can be written as \[P(\rho)d\rho = dV/V, \label{eq:PDFb}\] where \(dV/V\) is a fraction of volume with density in the range \((\rho,\rho+d\rho)\). The density dependance on radius can be inverted to give us the radius at a given density \(r=r(\rho)\). Then the PDF in eq.([\[eq:PDFb\]](#eq:PDFb){reference-type="ref" reference="eq:PDFb"}) takes the form \[P(\rho) =-\frac{4\pi r^2(\rho)}{V} \frac{dr(\rho)}{d\rho}. \label{eq:PDFc}\] This can be applied, for example, to the NFW profile. It is easy to see the trend, if the density profile is a power-law \(\rho\propto r^{-\alpha}\) with the slope \(\alpha\). In this case \[\rho^2P(\rho) \propto \rho^{1-3/\alpha}. \label{eq:PDFd}\] In the outer regions of a dark matter halo the density declines as a power-law with the slope \(\alpha\approx 3\), which implies that \(\rho^2P(\rho)\approx {\it constant}\). It is easy to invert the NFW profile numerically. Results show that for the halo masses in the range \(M=(10^{12}-10^{15})\mbox{\)h^{-1}M_\odot\( }\) the product \(\rho^2P(\rho)\) is nearly constant for overdensities \(\rho=10^2-10^4\) and declines at much larger densities. This decline is consistent with the fact that the slope \(\alpha\) becomes smaller at radii comparable with the characteristic scale radius \(r_s\) of the NFW profile. Adding results for many halos with different masses and concentrations will change the behavior of the \(\rho^2P(\rho)\) trend at \(\rho >10^4\), but not for the range \(\rho=10^2-10^4\), where it should remain flat because it is also flat for each halo. In summary, both the spherical infall model and the dark matter halo profiles indicate that the leading term in the density distribution function should be \(P(\rho)\propto\rho^{-2}\). The trend should be modified by adding suppression on large and small scales. However insightful, the spherical infall model or the NFW results at large densities cannot be used to produce accurate results for the density distribution function. We use those hints to construct our own approximation for the PDF \(P(\rho)\). Motivated by these results and by our simulations we design our own model. It is nearly a power-law \(P(\rho)\propto\rho^{-\alpha}\) with the slope \(\alpha \approx 2\) that is truncated with exponents on both the small and large densities. We call this model **double-exponential**. We tested different shapes for the exponential terms and find that the following expression provides errors less than few percent at all redshifts, smoothing scales, and cosmologies that we study: \[P(\rho) = A\rho^{-\alpha}\exp\left[-\left(\frac{\rho_0}{\rho}\right)^{1.1}\right] \exp\left[-\left(\frac{\rho}{\rho_1}\right)^{0.55}\right], \label{eq:dexp}\] where \(A,\, \alpha\), \(\rho_0\), and \(\rho_1\) are free parameters. As noticed above the slope \(\alpha \approx 2\). The slopes in the exponential terms 0.55 and 1.1 are results of the fitting of numerical PDFs at different smoothing scales and redshifts. One may expect that adding two more free parameters to the approximation (i.e., the slopes in the exponential terms) may further improve the quality of the fits. We find that this is not the case: the data prefer the same slopes regardless of the value of \(\sigma\). The double-exponential model has four formal free parameters. One may use three constraints to limit the parameters: the total mass and volume should be equal to unity (see eq. ([\[eq:normalization\]](#eq:normalization){reference-type="ref" reference="eq:normalization"})), and the second moment of the PDF should be equal to \(\sigma\) measured in simulations (see eq. ([\[eq:sigma\]](#eq:sigma){reference-type="ref" reference="eq:sigma"})). Note that there must be a degree of freedom left after fixing the constraints otherwise the model would not be able to reproduce the numerical results that show that the PDF is not defined solely by \(\sigma\), and depends on both the redshift and \(\Omega_m\). The double-exponential model has this additional degree of freedom. In practice, we use all four parameters to fit the numerical data. We typically find that the best-fit parameters provide PDF approximations that within 1-2% conserve the mass and match well the numerical value of \(\sigma\) measured in the simulations. The volume is conserved within 1-5% accuracy.
# Simulations {#sec:sim}
Numerical parameters of our simulations are presented in Table [\[table:simtable\]](#table:simtable){reference-type="ref" reference="table:simtable"}, which gives box size, number of particles, mass of a particle \(m_p\), number of mesh points \(N_g^3\) (if relevant), cell size of the density/force mesh \(\epsilon\), the number of time-steps \(N_s\), cosmological parameters \(\sigma_8\) and \(\Omega_m\), and number of realizations \(N_r\). In order to estimate the density distribution function, we split each simulation box with a \(3-D\) mesh with the cell size \(\Delta x\). We then use the Cloud-In-Cell (CIC) density assignment to generate the density field. Many filtering sizes \(\Delta x\) were used for each simulation and snapshot. Different codes were used to make those simulations. The MultiDark \(1~Gpc/h\) simulation (MDPL1) was done with the GADGET-2 code. The ART code was used to produce the BolshoiP simulation. These two simulations have the largest resolution and the largest number-density of particles. However, there are only two of these simulations because they are very expensive computationally. Other simulations were carried out with the parallel Particle-Mesh code GLAM. Because the GLAM code is much faster, we have many realizations of the same cosmological and numerical parameters. Simulations B0.5 and D0.5 were designed for testing possible dependance of the PDF on the matter density \(\Omega_m\). These simulations have the same random seeds to make comparisons of results easier. All the simulations were started at initial redshift \(z_{\rm init}=100\) using the Zeldovich approximation. The simulations span three orders of magnitude in mass resolution, a factor of hundred in force resolution and differ by a factor of 500 in effective volume (see Table [\[table:simtable\]](#table:simtable){reference-type="ref" reference="table:simtable"}). Altogether, we use about 7000 simulations to study the density distribution function. To our knowledge this is the largest set of simulations available today for this type of analysis. Analysis of different numerical effects in the estimates of the PDF is presented in Appendix A. In summary, our results are mostly dominated by systematics, not by the finite-volume simulation variance. When dealing with individual simulations such as MDPL1 or BolshoiP we use only bins with more than \(N>100\) cells per bin. For the large sets of simulations A0.5, A1.5, A2.5 we accept bins with more than 10 cells. The discreteness of density assignment may become an issue at small densities while the force resolution may affect the high-density tail of the PDF. We find limits on numerical parameters that should be satisfied to produce the \(P(\rho)\) with errors less than few percent: (1) the filtering scale \(\Delta x\) must be resolved with not less than 8 force resolution elements: \(\Delta x> 8\epsilon\), and (2) the number of particles per filtering cell should be not less than 10-20.
# PDF: main trends {#sec:main}
The overall dependance of \(P(\rho)\) on filtering scale and \(rms\) density fluctuation \(\sigma\) is illustrated in Figures [\[fig:pdf\]](#fig:pdf){reference-type="ref" reference="fig:pdf"} and [\[fig:pdf2\]](#fig:pdf2){reference-type="ref" reference="fig:pdf2"}. Full curves in the plots show the results from our simulations. At small \(\sigma\) and large \(\Delta x\) the PDF has a peak at \(\rho\approx 1\) that shifts to smaller densities as \(\sigma\) increases (and smoothing scale \(\Delta x\) decreases). At the same time the distribution function becomes extremely wide and develops a distinct \(\rho^{-2}\) power-law trend that expands both to large and small densities. By scaling out the \(\rho^{-2}\) dependance (Figures [\[fig:pdf2\]](#fig:pdf2){reference-type="ref" reference="fig:pdf2"}) we reduce the dynamical range and can see better details of the PDF. In particular, we find a very steep decline of \(\rho^{2}P(\rho)\) on both high and low density tails. The double-exponential model eq.([\[eq:dexp\]](#eq:dexp){reference-type="ref" reference="eq:dexp"}) was tuned to find the shape of both declines. Our results for different filtering scales and different redshifts show that the decline at the large-density limit is \(\propto \exp[-(\rho/\rho_1)^{\nu}]\) and at the small-density limit it is \(\propto \exp[-(\rho/\rho_0)^{-2\nu}]\) with \(\nu\approx 0.55\). Note that at very large \(\sigma \gtrsim 10\) the peak of \(\rho^{2}P(\rho)\) has a nearly constant amplitude \(\rho^{2}P(\rho)\approx 0.11\) but the position of the peak shifts to larger values of \(\rho\). Because the total mass must be preserved (\(\int \rho P(\rho)d\rho =1\)), this implies that at intermediate scales \(\rho\approx 1-100\) the PDF \(P(\rho)\) should decline when \(\sigma\) increases, and that the slope \(\alpha\) in eq.([\[eq:dexp\]](#eq:dexp){reference-type="ref" reference="eq:dexp"}) must become slightly shallower with increasing \(\sigma\). It is often taken for granted that the PDF depends only on the amplitude of the density perturbations on a given filtering scale \(\sigma\). Indeed, this is the dominant behavior of \(P(\rho)\). However, this is not exactly correct. Our simulations have such a good accuracy that now we can test the dependance of the PDF on redshift at fixed \(\sigma\) and on cosmological parameters. We first select redshifts and smoothing scales in such a way that \(\sigma\) for two different redshifts are nearly identical. Figure [\[fig:amplt\]](#fig:amplt){reference-type="ref" reference="fig:amplt"} presents two examples of such cases--one for relatively low \(\sigma\approx 1\) and another for larger \(\sigma\approx 4\). The differences between PDFs at the same \(\sigma\) and different \(z\) are not large: 5-10% depending on the density where the differences are measured. Nevertheless, the differences clearly exist. We also estimate the differences using the spherical infall model and present the results in the right panel of Figure [\[fig:amplt\]](#fig:amplt){reference-type="ref" reference="fig:amplt"}. The PDF also slightly depends on parameters of the cosmological model. In the left panel of Figure [\[fig:Omega\]](#fig:Omega){reference-type="ref" reference="fig:Omega"} we compare \(z=0\) results for the B0.5Gpc and D0.5Gpc simulations that differ only by the matter density parameter \(\Omega_m\). Again, differences are small but clearly present at \(\sim 5-10\%\) level. The right panel shows predictions for the spherical infall model with the same basic conclusion: PDF does depend on \(\Omega_m\). One can understand why the PDF depends on \(z\) and \(\Omega_m\), if one realizes that at any given density \(\rho\) the value of \(P(\rho)\) is formally a functional on the non-linear power spectrum. This means that it depends not only on \(\sigma(R)\) but also on the whole shape of the power spectrum. We can analyze the situation by assuming that \(P(\rho)\) depends just on \(\sigma_{L}(R)\) (see eq. ([\[eq:sigmaL\]](#eq:sigmaL){reference-type="ref" reference="eq:sigmaL"})) and on its local logarithmic derivative \(\gamma\) at scale \(R\) as defined by eq. ([\[eq:gam\]](#eq:gam){reference-type="ref" reference="eq:gam"}). Then the dependence of the PDF on redshift for a fixed value of \(\sigma_{L}(R)\) can be explained, because at larger redshits the given value of \(\sigma_{L}\) is attained at a smaller scale \(R\), where \(\sigma_{L}(R)\) is less steep for a CDM-type power spectrum. If we look at different terms for \(\rho^2P(\rho)\) in eq. ([\[eq:sph1\]](#eq:sph1){reference-type="ref" reference="eq:sph1"}), then we note that the \(\sigma_L\) in the argument of the exponential term is also a function of \(\rho\) (falling faster with \(\rho\) as \(\gamma\) grows). So, for larger values of \(\gamma\) the PDF will be smaller for large \(\rho\) values, where the exponential behavior dominates, while for small values of \(\rho\) it will be larger. The derivative in the right-hand-side of eq. ([\[eq:sph1\]](#eq:sph1){reference-type="ref" reference="eq:sph1"}) also depends on \(\gamma\): it will be larger for larger \(\gamma\) values. However, it is only important for the intermediate values of density, where the behavior can qualitatively be explained by the behavior in the extremes and the conservation of probability. In short, this all implies a smaller PDF in the low \(\rho\) limit and a larger one in the large \(\rho\) limit as compared to the PDF at smaller redshift. The same qualitative behavior would be observed for smaller values of \(\Omega_m\) when studying the dependence of the PDF on \(\Omega_m\) at a given redshift: as \(\Omega_m\) decreases the power spectrum flattens (up to scales of the order of the horizon at the start of matter domination), leading to a less steep \(\sigma_L.\)
# PDF: Testing different approximations and models. {#sec:sphLog}
We start our analysis of different approximations by testing the spherical infall model and the log-normal distribution. Both models are expected to work and typically used for relatively low rms fluctuations \(\sigma\lesssim 1\). Figure [\[fig:spher\]](#fig:spher){reference-type="ref" reference="fig:spher"} shows results for two modifications of the spherical infall model. We select five configurations with different filtering scales, \(\sigma\) and redshifts. There are some differences between the pure spherical infall model in eq. ([\[eq:sph1\]](#eq:sph1){reference-type="ref" reference="eq:sph1"}) and the excursion set model in eq. ([\[eq:sph2\]](#eq:sph2){reference-type="ref" reference="eq:sph2"}). For example, the excursion set model produces smaller errors at \(\rho<1\) and \(\sigma\approx 1\). At the same time, it makes visibly larger errors at \(\rho>1\). For this reason we prefer the standard spherical infall model. It provides smaller than 10% errors for points that are larger than 0.1 of the maximum of the PDF. The error increases substantially in the peripheral regions. If better accuracy is required for the small values of the PDF, one would need to use results of \(N\)-body simulations, not the approximations. Results for the log-normal distribution are shown in Figure [\[fig:lognorm\]](#fig:lognorm){reference-type="ref" reference="fig:lognorm"}, where in the left panel we present \(z=0\) results, and results for different redshifts are shown in the right panel. It is clear that the log-normal distribution produces much worse fits as compared with the spherical infall model. For example, errors in the central region are less than 10% only for \(\sigma<0.5\). They become dramatically worse for even slightly larger \(\sigma\). For \(\sigma>1\) the lognormal distribution makes typically \(\sim 50\%\) errors and, which is even worse, predicts a wrong shape of the PDF. It predicts a wrong position of the maximum; slopes of the declining PDF are not correct. The log-normal distribution has two advantages as compared with the spherical infall model: (a) It is simple; does not require the machinery of handling the power spectrum and numerical derivatives. (b) it makes mediocre predictions that do not totally fail. We definitely recommend it for quick-and-dirty applications, but not for accurate estimates. Left panels in Figure [\[fig:NBN\]](#fig:NBN){reference-type="ref" reference="fig:NBN"} show results for the Negative Binomial distribution given by eq. ([\[eq:nbn\]](#eq:nbn){reference-type="ref" reference="eq:nbn"}). For the large smoothing scale \(\Delta x=20\mbox{\)h^{-1}\(Mpc}\) it provides a reasonable fit with errors \(\sim 10\%\) for densities \(\rho =0.4-2.5\). However, it is clear that it has a wrong shape: too steep at large densities and too shallow at small densities. This becomes a serious issue for large values of \(\sigma\). For example, we could not find any good fit for \(\Delta x=5\mbox{\)h^{-1}\(Mpc}\) shown in the bottom panel. It may be not fare to use the NBN for the dark matter PDF because we are in a regime that was not favorable for the NBN: study configurations with very large number of particles per cell while the NBN was designed to handle very small number \(\bar N\). The GEV approximation eq. ([\[eq:GEV\]](#eq:GEV){reference-type="ref" reference="eq:GEV"}) scored much better, as illustrated in the right panels of Figure [\[fig:NBN\]](#fig:NBN){reference-type="ref" reference="fig:NBN"}. Indeed, it provided excellent fits for \(\sigma \lesssim 2\) with errors less than 10% and even for larger \(\sigma\) it gives very good accuracy, but it starts to fail catastrophically at very large densities. tested this approximation for PDF for the Bolshoi simulation. Their results are compatible with ours. However, seems did not pay attention to the situation at large densities, where the GEV becomes unacceptable. It still can be useful for low densities, but the problem is that there is no a priory estimate at what density and \(\sigma\) the GEV fails. Comparison of the double-exponential model with the \(N\)-body results has been already presented in Figure [\[fig:pdf\]](#fig:pdf){reference-type="ref" reference="fig:pdf"}. More detailed analysis is shown in Figure [\[fig:doubleexp\]](#fig:doubleexp){reference-type="ref" reference="fig:doubleexp"}. Parameters of the fits for different smoothing scales and redshifts are given in Appendix B. Of all approximations studied in this paper the double exponential model is by far the best. It provides very accurate (few percent) fits for \(\sigma \gtrsim 1\) with densities \(\rho\approx 10^{-1}-10^5\). With somewhat larger errors it still works down to \(\sigma \approx 0.5\). To some degree the success of the approximation is not surprising because it was designed to reproduce the main features of the spherical infall model and the PDF of individual NFW halo profiles that predict the \(P(\rho)\propto \rho^{-2}\) trend for large densities. And, of course, the parameters of the approximation--the two power-law slopes--were tuned to produce best fits. However, it is unexpected that the exponential terms in eq. ([\[eq:dexp\]](#eq:dexp){reference-type="ref" reference="eq:dexp"}) should have the same power-law slopes 0.55 and 1.11 for all densities and smoothing scales. Indeed, we tried different combinations of the slopes--even with changing values of the slopes for different \(\sigma\)--and did not find them to improve the fits. There is one problem with the constant slopes, though: the approximation cannot work for a very small \(\sigma\) where the PDF must become a Gaussian. The eq. ([\[eq:dexp\]](#eq:dexp){reference-type="ref" reference="eq:dexp"}) does not allow a transition to a Gaussian distribution. This is not a serious issue because at \(\sigma \lesssim 1\) the spherical infall model provides an adequate approximation for the dark matter density distribution function.
# Conclusions {#sec:concl}
Using a large suite of cosmological \(N\)-body simulations we study the shape and the evolution of the dark matter one-point probability distribution function. Unlike most of other studies, we cover a very large range of smoothing scales \(R=100\,\mbox{\)h^{-1}\(kpc}-20\,\mbox{\)h^{-1}\(Mpc}\) and \(rms\) density fluctuations \(\sigma\). We find that as \(\sigma\) increases, the PDF becomes a power-law \(P\propto\rho^{-2}\) that is truncated with exponential terms on both small and large densities. This trend is consistent with the extrapolation of both the spherical infall model and the PDF expected at high densities for the NFW density profile of dark matter halos. The PDF weakly depends on redshift (at fixed \(\sigma\)) and on matter density \(\Omega_m\). The effect is relatively small (\(\sim 5-10\%\)) but is clearly observed in simulations. The spherical infall model also has the same trend. This behavior contradicts analytical approximations such as the log-normal or the GEV that assume that the PDF should depend only on \(\sigma\). The basic trend \(P\propto\rho^{-2}\) gives us a motivation to construct a new model given in eq. ([\[eq:dexp\]](#eq:dexp){reference-type="ref" reference="eq:dexp"}), which we call double-exponential distribution. It formally has four free parameters of which two can be fixed by requiring that the total volume and mass must be equal to unity. The model works only for \(\sigma\gtrsim 1\) and does not allow the transition to a Gaussian distribution as expected for \(\sigma\ll 1\). Nevertheless, for \(\sigma\gtrsim 1\) the model gives the best performance of all approximations that we tested in this work, with errors of just few percent for \(P(\rho)\) when the PDF changes by 12 orders of magnitude. Parameters of the double-exponential model are provided in Appendix B. The model potentially may be modified by adding few extra parameters to allow for accurate treatment at small \(\sigma\). We did not try to do it for two reasons: (1) the spherical infall model gives accurate enough treatment for this regime, and (2) it is cheap to make adequately accurate \(N\)-body simulations for \(\sigma<1\) if needed. The spherical infall model provides accurate predictions for the \(N\)-body PDF results for low values of \(\sigma<1\), but it becomes less reliable for larger \(\sigma\), which is expected for this model. The combination of the double-exponential model at large \(rms\) fluctuation with the spherical infall model in the small \(rms\) regime yields a remarkably accurate density distribution for all regimes of clustering of the cosmological matter field. We also tested different analytical approximations. The often used log-normal distribution, as was shown before, does not make good fits and needs significant modifications before it can provide accurate results. It is pretty much useless for large densities because it does not provide a path to explain the main trend at large densities, i.e. the power-law trend \(P\propto\rho^{-2}\). It clearly has an advantage of being simple, and it does not fail catastrophically as shown in Figure [\[fig:lognorm\]](#fig:lognorm){reference-type="ref" reference="fig:lognorm"}. Our results for \(\sigma<1\) demonstrate that the log-normal approximation always made significantly worse fits as compared with the spherical infall model predictions. In addition, the log-normal approximation cannot accomodate the dependance of the PDF on redshift and \(\Omega_m\), while the spherical infall model nicely reproduces the effect. The GEV approximation scores much better than the log-normal distribution. Even for a large \(\sigma\approx 10\) it gives remarkably accurate results for densities up to \(\rho\approx 10^3\). However, the approximation fails catastrophically at larger densities, and there is no obvious way to predict at what density it still works or fails. While it is useful and insightful to have analytical models for the PDF, one does not really need them if \(\sim 1\)% accuracy is a requirement and \(\sigma\gtrsim 0.5\). Then there is no alternative to \(N\)-body simulations: one can get very accurate, fast and cheap results. The cost of one GLAM A0.5 simulation is just 3.5 hrs of wall clock on a modest data server. For the B0.5 run it is 10 hrs. In this paper we analyse thousands of realizations, but this was to prove that one does not need those to make accurate PDF: just a few realizations is enough. However, one needs to be carefull about numerical effects when using these fast Particle-Mesh simulations. provide detailed description of constraints for the simulations, and Appendix A in this paper gives prescriptions on how to use the \(N\)-body simulations to reach \(\sim 1\%\) accuracy in the PDF. | {'timestamp': '2017-06-08T02:00:54', 'yymm': '1706', 'arxiv_id': '1706.01909', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01909'} |
null | null |
# Introduction {#sec:intro}
Electron spin resonance (ESR) and its variants in magnetically ordered systems-ferromagnetic and antiferromagnetic resonances-represent one of the most precise and frequently used spectroscopic probes of excitations of magnetic media. The essence of the magnetic resonance technique consists in measuring absorption of electromagnetic radiation (usually in the microwave range of frequencies) by a sample material which is (typically) subjected to an external static magnetic field. The absorption is caused by coupling of magnetic degrees of freedom to the magnetic field of the electromagnetic wave. Given very large wavelength of the microwave, the ESR absorption is driven by zero wavevector (\({\bf q} = 0\), or vertical) transitions between states with different \(S^z\) projections of magnetic dipole moment on the direction perpendicular to the magnetic field of the EM radiation. In a spin system with isotropic exchange, the absorption spectrum of an AC magnetic field is a \(\delta\)-function peak at the frequency equal to that of the Zeeman energy, independently of the exchange interaction strength. This is a consequence of the fact that at \({\bf q} = 0\) EM radiation couples to the total magnetic moment, which for an \(SU(2)\) invariant system commutes with the Hamiltonian. Therefore, any deviation of the absorption spectrum from the \(\delta\)-function shape implies violation of the spin-rotation symmetry, caused either by anisotropic terms in the Hamiltonian (explicit symmetry breaking) or by the development of long-range magnetic order below the critical temperature (spontaneous symmetry breaking). This is the key reason for ESR's utility. The goal of our work is to explore applications of ESR to highly entangled phase of magnetic matter-the quantum spin liquid (QSL). This intriguing novel quantum state manifests itself via non-local elementary excitations-spinons-which behave as fractions of ordinary spin waves. Local spin operator becomes a composite of two or more spinons, which immediately implies that dynamic spin susceptibility measures multi-spinon continuum. In principle, the best probe of the spinon continuum is provided by inelastic neutron scattering which probes spinons at finite wave vector \({\bf q}\) and frequency \(\omega\). By now several textbook-quality experiments have provided us with unambiguous signatures of multi-particle continua. In practice, however, such state of the art measurements require large high-quality single crystals which quite frequently are not available. We posit here that ESR, with its exceptionally high energy resolution, represents an appealing complimentary spectroscopic probe of spinons-*spinon magnetic resonance* (SMR). The key requirement for turning it into a full-fledged probe of spinon dynamics consists in the *absence of spin-rotational invariance*. This requirement stems from the mentioned above 'insensitivity' of ESR to the details of excitations spectra in \(SU(2)\) invariant magnetic materials. Note that the \(SU(2)\) invariance is, at best, a theoretical approximation to the real world materials which *always* suffer from some kind of magnetic anisotropy. Moreover, over the past fifteen years the field of QSL has evolved dramatically *away* from the spin-rotational invariance requirement explicit in many foundational papers. The absence of spin-rotational invariance has evolved from the 'real world' annoyance to the virtue. Indeed, the first and still the most direct and unambiguous demonstration of the gapless QSL phase came from Kitaev's exact solution of the fully anisotropic honeycomb lattice model which does not conserve total spin. Importantly, a large number of very interesting and not yet understood materials, such as \(\alpha\)-RuCl\(_3\), YbMgGaO\(_4\), Yb\(_2\)Ti\(_2\)O\(_7\) and many other pyrochlores, and even organic BEDT-TTF and BEDT-TSF salts, showing promising QSL-like features are known to possess significant spin-orbit interaction and are described by spin Hamiltonians with significant asymmetric exchange and pseudo-dipolar terms. It is precisely this class of low-symmetry spin models we focus on in the present study. We illustrate our idea by considering spin-liquid state proposed to describe a spin-orbit-coupled triangular lattice Mott insulator YbMgGaO\(_4\). The appropriate spin Hamiltonian has been argued to be that of XXZ model with interactions between nearest (with \(J \sim 1\)K) and next-nearest neighbors on the triangular lattice together with a pseudo-dipolar term, of \(J_{\pm\pm}\) kind in notations of, between nearest neighbors (\(J_{\pm\pm} \sim 0.2\)K). Most recently, polarized neutron scattering data were interpreted in favor of significant \(J_{z\pm}\) interaction. This Hamiltonian does not conserve total spin \({\bf S}_{\rm tot}\). Inelastic neutron scattering experiments reveal broad spin excitations continuum, consistent with fractionalized QSL with spinon Fermi surface. At the same time experimental evidence of significant disorder effects, capable of masking 'pristine' physics of the material, is mounting. Our goal here is to add to the ongoing discussion on the nature of the ground state of YbMgGaO\(_4\) by pointing out that ESR can serve as a very useful probe of QSL with significant built-in spin-orbit interactions. We therefore accept spin-liquid hypothesis and focus on fermionic \(U(1)\) symmetric spin-liquid ground states, proposed for this material previously. We rely on the well-established projective symmetry group (PSG) analysis of possible \(U(1)\) spin liquids. The spin-orbital nature of the effective spin-1/2 local moment of Yb\(^{3+}\) ion implies that under the space group symmetry operations both the direction and the position of the local spin are transformed. The symmetry operations include translations \(T_{1,2}\) along the major axis \({\bf a}_{1,2}\) of the crystal lattice, a rotation \(C_2\) by \(\pi\) around the in-plane vector \({\bf a}_1 + {\bf a}_2\), a counterclockwise rotation \(C_3\) by \(2\pi/3\) around the lattice site, and the (three-dimensional) inversion \(I\) about the lattice site. Following, it is convenient to combine \(C_3\) and \(I\) operations into a composite one \({\bar C}_6 \equiv C_3^{-1} I\). (Note that the original \(C_6\) lattice rotation by \(2\pi/6\) around the lattice site is *not* the symmetry of YbMgGaO\(_4\) due to alternating-above and below the plane-location of oxygens at the centers of consecutive elementary triangles.) These symmetries strongly constrain possible \(U(1)\) mean-field spinon Hamiltonians and result in 8 different PSG states, of U1A and U1B kind. U1A states maintain periodicity of the original lattice and their band structure consists of just two spinon bands. U1B states are \(\pi\)-flux states with doubled unit cell. Equivalently, their band structure contains 4 spinon bands. For the sake of simplicity, we focus on the U1A family in the following (description of U1B increases algebraic complexity without adding any new essential physics). The \(U(1)\) mean-field spinon Hamiltonian is parameterized by several hopping amplitudes-\(t_{1,2}\) describe spin-conserving hopping between the nearest and the next-nearest neighbors and \(t_{1,2}'\) describes analogous non-spin-conserving hops. PSG analysis fixes relative phases between hopping amplitudes on the bonds related by the space group operations (see Supplement for details of the derivation). The magnitudes of these hoppings are not determined by PSG. This requires a separate variational calculation of the ground state energy which is not attempted here. We do expect, on physical grounds, that for the spin model with predominant isotropic nearest-neighbor spin exchange and subleading asymmetric \(J_{\pm\pm}\) terms, the following estimate should hold \(t_1 > t_1' > t_2 > t_2'\). There are four mean-field Hamiltonians in U1A family, labeled by U1A\(n_{C_2}n_{\bar{C}_6}\) (\(n_{C_2},n_{\bar{C}_6}\in\{0,1\}\)).They have the simple form \[H = \sum_{\bf k} (f^\dagger_{{\bf k} \uparrow}, f^\dagger_{{\bf k} \downarrow}) \left(\begin{matrix} \omega_{\bf k} + \epsilon_{\bf k} & \eta_{\bf k} \\ \eta_{\bf k}^* & \omega_{\bf k}-\epsilon_{\bf k}\end{matrix}\right) \left(\begin{matrix} f_{{\bf k} \uparrow} \\ f_{{\bf k} \downarrow} \end{matrix}\right), \label{eq:1}\] where \({\bf k}\)-dependent \(\omega_{\bf k}, \epsilon_{\bf k}, \eta_{\bf k}\) are listed in. Spin-orbit interaction appears via spin-non-conserving hopping \(\eta_k\) in [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"}. The U1A00 state is characterized by finite \(\omega_{\bf k}\) and zero \(\epsilon_{\bf k}\) and \(\eta_{\bf k}\), while U1A01 and U1A11 have \(\omega_{\bf k} = 0\) and finite \(\epsilon_{\bf k}\) and \(\eta_{\bf k}\). In the calculations below we set \(t_1 =1\) and \(t_1' = 0.8, t_2' = 0.3\) for U1A11, while \(t_1 =0\) for U1A01 and we choose \(t_1' = 1, t_2 = 0.8 t_1', t_2' = 0.4 t_1'\) for it. U1A10 turns out to be non-physical since its Hamiltonian matrix is zero, \(t_{1,2} = t_{1,2}'=0\). 'Accidental' nature of U1A00 state is manifested by the absence of any spin-dependent hopping in its Hamiltonian-this state happens to be more symmetric than the spin Hamiltonian it describes and is characterized by the large Fermi surface. We focus on most physically relevant U1A01 and U1A11 states, for which \(\omega_{\bf k}=0\). The resulting fermion bands are easy to find, \(E_{\nu = 1,2}({\bf k}) = (-1)^\nu E({\bf k}) = (-1)^\nu \sqrt{\epsilon_{\bf k}^2 + |\eta_{\bf k}|^2}\). U1A11 state possess symmetry-protected Dirac nodes at \(\Gamma\) and M points of the hexagonal Brillouin zone, while U1A01 has additional Dirac nodes at K points as well. Interaction with monochromatic radiation linearly polarized along direction \(\hat{{\bf n}} = (\sin\theta \cos\phi, \sin\theta \sin\phi, \cos\theta)\) is described by \(V(t) =-{\bf h}(t) \cdot {\bf S}_{\rm tot}\), *i.e.* \[V(t) = h e^{-i\omega t} {\bf n} \cdot \frac{1}{2}\sum_{\bf r} (f^\dagger_{{\bf r} \uparrow}, f^\dagger_{{\bf r} \downarrow}) {\bm \sigma} \left(\begin{matrix} f_{{\bf r} \uparrow} \\ f_{{\bf r} \downarrow} \end{matrix}\right) \label{eq:2}\] Within linear response theory the rate of energy absorption \(I(\omega) =-\omega \chi''_{\rm n n}(\omega) |h|^2/2\) is determined by the imaginary part of \({\bf q}=0\) Fourier transform of the dynamic susceptibility \(\chi_{\rm n n}(t, {\bf r}) =-i \Theta(t) \langle [{\bf S}_{\bf r}(t) \cdot \hat{{\bf n}}, {\bf S}_{\bf 0}(0) \cdot \hat{{\bf n}}]\rangle\), with \(\Theta\) being the Heaviside function. Straightforward calculation gives \[\begin{aligned}
\chi_{\rm n n}(\omega) &=& \frac{1}{4N}\sum_{\bf k} \frac{n_{{\bf k} \alpha}-n_{{\bf k} \beta}}{\omega + E_\alpha({\bf k})-E_\beta({\bf k}) + i0} \nonumber\\ &&\times (U_{\bf k}^+ \sigma^a U_{\bf k})_{\alpha \beta} (U_{\bf k}^+ \sigma^b U_{\bf k})_{\beta \alpha} \hat{n}^a \hat{n}^b \label{eq:3} \end{aligned}\] Here \(n_{{\bf k} \alpha}\) is the occupation number of the band \(\alpha\), \(U_{\bf k}\) is unitary diagonalizing matrix connecting spinor of original fermions to that of the band ones, \(f_{{\bf k}, \alpha} = (U_{{\bf k}})_{\alpha\beta} b_{{\bf k},\beta}\), and summation over repeated indices is implied. Eq.[\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} shows that in the spin-degenerate U1A00 state, for which \(n_{{\bf k} \alpha} = n_{{\bf k} \beta}\), the susceptibility is strictly zero. Therefore, in agreement with general discussion above, no energy absorption occurs in the absence of external magnetic field for this state. The condition \(n_{{\bf k} \alpha} \approx n_{{\bf k} \beta}\) is also satisfied at high temperature of the order of spinon bandwidth (which is of the order of exchange \(J\)) when spinon resonance disappears. We therefore expect the width of the resonance to *increase* when the temperature is lowered. It is worth noting that the lowest temperature of ESR study is \(1.8\)K, which makes it a high-temperature measurement. At zero temperature absorption at frequency \(\omega\) is possible only via vertical transitions from the filled lower band (\(\alpha = 1, n_{{\bf k} 1}=1\)) to the empty upper one (\(\beta = 2, n_{{\bf k} 2}=0\)) and therefore \[\begin{aligned}
\chi''_{\rm n n}(\omega) &=&-\frac{\pi}{4N}\sum_{\bf k} \delta(\omega-2 E({\bf k})) (U_{\bf k}^+ \sigma^a U_{\bf k})_{12} \nonumber\\ && \times (U_{\bf k}^+ \sigma^b U_{\bf k})_{21} \hat{n}^a \hat{n}^b \label{eq:4} \end{aligned}\] After some algebra, the product of matrix elements \(12\) and \(21\) of the rotated Pauli matrices in the equation above simplifies to \[\begin{aligned}
\chi''_{\rm n n}(\omega) &=&-\frac{\pi}{4N}\sum_{\bf k} \frac{\delta(\omega-2 E({\bf k}))}{E({\bf k})^2} \left[(\epsilon_{\bf k}^2 + \eta_{\bf k}''^2) \sin^2\theta\cos^2\phi\right. \nonumber\\ &&\left.+(\epsilon_{\bf k}^2 + \eta_{\bf k}'^2) \sin^2\theta\sin^2\phi + |\eta_{\bf k}|^2\cos^2\theta \right]. \label{eq:5} \end{aligned}\] It can be shown that the omitted off-diagonal terms, containing products \(\hat{n}^x \hat{n}^y, \hat{n}^x \hat{n}^z\) and \(\hat{n}^y \hat{n}^z\), are all zero. Moreover, terms proportional to \(\cos^2\phi\) and \(\sin^2\phi\) are actually equal, so that the absorption only depends on the azimuthal angle \(\theta\) with respect to the normal to the magnetic layer. Two features of this result are worth noting. First, the absorption takes place over the wide band of frequencies, \({\rm min}(E) = 0 < \omega/2 < {\rm max}(E)\), which covers the full bandwidth of two-spinon continuum. Second, Eq.[\[eq:5\]](#eq:5){reference-type="eqref" reference="eq:5"} describes *zero-field absorption*, which does not require any external static magnetic field \({\bf B}\). Both of these are direct consequence of the absence of spin conservation in Eq.[\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"}. **U1A11 state:** Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} shows scaled absorption intensity, \(2 I(\omega)/|h|^2 =-\omega \chi''_{\rm nn}(\omega)\), for different polarizations. Polarization dependence is strong. The plot is obtained by numerical integration of [\[eq:5\]](#eq:5){reference-type="eqref" reference="eq:5"}, with frequency steps of \(\Delta \omega=0.05\), over the primitive cell of the reciprocal lattice (\({\bf k} = (k_1,k_2)\), where \(k_{1,2} \in (0,2\pi)\), see for details). We approximate the delta-function by the Lorenzian \(\delta(x) \approx \pi^{-1} d/(d^2 + x^2)\) with \(d=0.01\). We checked that \(d=0.05\) results in the same outcome. As expected, and also easy to check analytically, \(\chi''_{\rm nn}(\omega) \sim \omega\) at small frequencies. This is the consequence of Dirac nodes at \(\Gamma\) and M points. Behavior near the upper boundary, \(\omega \approx 3\sqrt{3}\), is determined by the vicinity of K point where \(\epsilon({\bf K}) = {\text{const}}\) while \(\eta({\bf K}) = 0\). As a result, one obtains \(\chi''_{\rm zz} \sim 3\sqrt{3}-\omega\), while at \(\theta = \pi/2\) susceptibility terminates discontinuously in a step-like fashion, \(\chi''_{\rm nn} = \chi''_{\rm xx} \sim \Theta(3\sqrt{3}-\omega)\). The rounding of the step-function behavior in Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}, for \(\theta \neq 0\), is caused by the finite width of numerical delta-function used in the integration over the Brillouin zone. The peak in the middle of the absorption band, at \(\omega \approx 2.43\), is caused by the van Hove singularity, of the saddle point kind, of \(E({\bf k})\) at \({\bf k}_0 = (k_0, 2\pi-k_0)\) and symmetry-related points. Here \(k_0 \approx 1.97\) and \(E({\bf k}\approx {\bf k}_0) \approx 1.215 + 1.31 (k_1 + k_2)^2-0.23 (k_1-k_2)^2\). The saddle-point produces logarithmically divergent contribution, \(\chi''_{\rm nn} \sim \ln|\omega-2.43|\), which matches numerical data in Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} perfectly. **U1A01 state:** SMR of this phase is shown in Figure [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}. It is seen to host two van Hove singularities which can be qualitatively understood as a direct consequence of the additional, in comparison with U1A11 state, Dirac cone in the spinon dispersion at the K point. The presence of the symmetry-protected node at the K point results in a stronger variation of spinon dispersion in the Brillouin zone and causes the appearance of additional saddle points. **U1A11 state in magnetic field:** external magnetic field adds further variations to the spinon absorption intensity. We illustrate this with the case of U1A11 state subject to magnetic field \({\bf B} = B_z \hat{z}\) along the normal to the magnetic layer. It should be noted that PSG analysis underlaying our consideration assumes time-reversal (TR) symmetry. Therefore we treat magnetic field perturbatively, by coupling it to the local TR-odd combination of spinons which is just \(B_z S^z_{\bf r} \sim B_z f^\dagger_{{\bf r} \alpha} \sigma^z_{\alpha \beta} f_{{\bf r} \beta}\). Thus magnetic field enters [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"} via \(\epsilon_{\bf k} \to \epsilon_{\bf k}-B_z/2\) and gaps out Dirac nodes. The minimal excitation energy becomes \({\text{min}}(E) = B_z/2\) and absorption intensity acquires threshold behavior \(I(\omega) \sim \Theta(\omega-B_z)\). This behavior is illustrated in Figure [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}, which also shows development of additional spectral features at \(\omega \approx 4.2\), see. In-plane magnetic field lowers symmetry of the spin Hamiltonian further and its consideration is left for future studies. This unusual response should be contrasted with that of the large-Fermi-surface state U1A00. Here \({\bf B} = B_z \hat{z}\) leads to the Zeeman splitting of spinon up-and down-spin bands \(E_\nu = \omega_{\bf k} \mp B_z/2\) and therefore, according to [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}, one finds standard result for magnetically isotropic media \(\chi''_{\rm n n}(\omega) \sim \sin^2\theta ~\delta(\omega-B_z)\). This is consistent with the earlier analysis of, where a weak magnetic field \({\bf B} = B_z \hat{z}\) was added to the mean-field Hamiltonian similarly. Off the \(\Gamma\) point, i.e. for \({\bf q}\neq 0\), one finds broad continuum corresponding to the spinon particle-hole excitations. **Discussion:** Physical arguments leading to Eq.[\[eq:5\]](#eq:5){reference-type="eqref" reference="eq:5"} are very general and rely on absence of long-range magnetic order, existence of fractionalized elementary excitations, which ensure a continuum-like response to external probes, and significant built-in spin-orbit interaction, which leads to non-conservation of spin and makes zero-field absorption possible in a wide range of frequencies. All of these are very generic conditions which are satisfied by essentially every model of spin liquids of \(U(1)\) and \(Z_2\) type (but not by spin-conserving \(SU(2)\) ones). The restriction to low symmetry spin liquids is not really a handicap as it turned out that the number of possible spin liquids with reduced \(U(1)\) and \(Z_2\) vastly outnumbers that of \(SU(2)\) symmetric ones. In particular, the SMR should be present in the celebrated Kitaev's honeycomb model, as was emphasized in dynamic structure calculations of. There too one can see anisotropic spin structure factor \({\cal S}^{aa}({\bf q}=0,\omega)\), with \({\cal S}^{zz} \neq {\cal S}^{xx/yy}\), and sharp van Hove singularities in the Majorana fermion density of states. The similarity is not accidental-it follows from the linear mapping between Majorana and projective spinon representations. Unlike the situation described here, in the exactly solvable gapless Abelian region dynamic response appears above a finite threshold energy (which is the energy cost of creating \(Z_2\) fluxes). However generic spin exchange perturbations turn the response gapless, so that \({\cal S}^{aa}({\bf q}=0,\omega) \sim \omega\) at low energy. Resonant inelastic x-ray (RIXS), Raman scattering and parametric pumping of the \(Z_2\) Kitaev spin liquid results in a gapless and extended in energy continuum too. Our theory can be broadly thought of as an extension of one-dimensional theories of ESR in spin chains with Dzyaloshinskii-Moriya interactions . In one dimension, fractionalized nature of spinons is very well established and theories based on them describe ESR experiments exceedingly well, both in gapless and gapped settings. Another important connection is provided by electric dipole spin resonance (EDSR) which describes absorption of EM radiation in *conductors* with pronounced spin-orbit interaction which mediates coupling of AC electric field to the electron spin. Here, spin-rotational asymmetry causes strong absorption which is controlled by the real part of optical conductivity. Somewhat surprisingly, energy absorption due to coupling of spins to AC electric field is also possible in strong Mott insulators, provided they are built of frustrated triangular units, in which virtual charge fluctuations produce spin-dependent electric polarization . Hints of this physics were recently observed in herbertsmithite and \(\alpha\)-RuCl\(_3\) antiferromagnets. Simple calculations of SMR presented here are based on mean-field spinon Hamiltonians derived with the help of PSG formalism. They do not include gauge fluctuations which undoubtedly are present in the theory. These fluctuations are certain to affect exponents characterizing sharp features of \(\chi''_{\rm n n}(\omega)\), such as for example behavior near the van Hove singularity and/or near lower/upper edge of the two-spinon continuum. (Disorder, in the form of Mg/Ga mixing, leads to distribution of g-factors which also broadens magnetic response.) In addition, by analogy with critical Heisenberg chain, we expect four-spinon contributions to the susceptibility to affect the high-frequency behavior. However, these important effects *can not reduce* spinon absorption bandwidth and eliminate other outstanding features of the SMR found here. It should also be noted that SMR is not specific to fermionic spinons and indeed extension of the theory to bosonic PSG is possible as well. We therefore conclude that spinon magnetic resonance represents an efficient and informative probe of exotic excitations of spin-orbit-coupled quantum spin liquids.
**Notes Added.** Manuscript, which appeared after our submission, contains detailed comparison of ground state energies of various U(1) PSG states. | {'timestamp': '2018-01-23T02:11:26', 'yymm': '1706', 'arxiv_id': '1706.01597', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01597'} |
null | null |
# Introduction
Given two smooth projective varieties \(X\) and \(Y\) related by a Fourier-Mukai transform \(\Phi: D^b(X) \to D^b(Y)\), it is usually not the case that a slope stable coherent sheaf \(E\) on \(X\) is taken by \(\Phi\) to a slope stable coherent sheaf on \(Y\). It is therefore natural to ask: under what circumstances does this happen? Yoshioka gave examples and counterexamples to the above question on Abelian and K3 surfaces. On threefolds, Bridgeland-Maciocia showed in that, if \(X\) and \(Y\) are dual elliptic threefolds with a corresponding Fourier-Mukai transform \(\Phi: D^b(X) \to D^b(Y)\), then for any rank-one torsion-free sheaf \(E\) on \(X\), there is a fixed line bundle \(L\) (depending only on \(\ch_1(E)\)) such that \(E \otimes L\) is taken by \(\Phi\) to a torsion-free sheaf that is Gieseker stable with respect to a polarisation \(\omega\) on \(Y\), where \(\omega\) depends on \(\ch(E)\). The main result of this paper is as follows: **Theorem [\[theorem:main1\]](#theorem:main1){reference-type="ref" reference="theorem:main1"}**. *Suppose \(p: X \to S\) is a Weierstraß threefold where \(X\) is \(K\)-trivial and \(K_S\) is numerically trivial. Then for any ample class \(\omega\) on \(X\) and any line bundle \(M\) on \(X\) of nonzero fiber degree, the Fourier-Mukai transform of \(M\) (up to a shift) is a \(\mu_\omega\)-stable locally free sheaf.* A main ingredient in the proof of Theorem [\[theorem:main1\]](#theorem:main1){reference-type="ref" reference="theorem:main1"} is a pair of spectral sequences on the interplay between a Fourier-Mukai transform and the derived dual functor on elliptic threefolds, discussed in Section [3](#sec:specseq){reference-type="ref" reference="sec:specseq"}. These spectral sequences are inspired by analogues on Abelian threefolds due to Maciocia-Piyaratne. We note that some of the techniques in this paper have also appeared in the work of Oberdieck-Shen. In their work, they study the transform of stable pairs (in the sense of Pandharipande-Thomas ) under an autoequivalence of the derived category of an elliptic threefold, thereby giving a partial proof to the modulairty conjecture on PT invariants due to Huang-Katz-Klemm. The paper is organized as follows: in Section [2](#sec2){reference-type="ref" reference="sec2"} we recall the Fourier-Mukai functor for Weierstraß fibrations and the dualizing functor, as well as Theorem [\[thm:commutativity\]](#thm:commutativity){reference-type="ref" reference="thm:commutativity"} that establishes the commutativity of the two functors. In Section [3](#sec:specseq){reference-type="ref" reference="sec:specseq"} we apply Theorem [\[thm:commutativity\]](#thm:commutativity){reference-type="ref" reference="thm:commutativity"} and the technique of spectral sequences to study the behavior of sheaves under Fourier-Mukai transforms. In Section [4](#sec4){reference-type="ref" reference="sec4"}, we give a criterion for reflexive sheaves to be taken to locally free sheaves by a Fourier-Mukai transform, thus paving the way for the proof of the main result of this paper, Theorem [\[theorem:main1\]](#theorem:main1){reference-type="ref" reference="theorem:main1"}. **Acknowledgements.** We would like to thank Arend Bayer, Antony Maciocia and Dulip Piyaratne for valuable discussions that led to the conception of this project. The first author would also like to thank the Institute of Algebraic Geometry at Leibniz University Hannover for their hospitality and support during the final stage of this project in May-June 2017. The second author would like to thank Alastair Craw for his support via EPSRC grant EP/J019410/1 and the support from Leibniz University Hannover during various stages of the work.
# Commutativity of Fourier-Mukai and dualizing functors {#sec2}
In this section, all functors are derived unless otherwise specified.
## The Fourier-Mukai functor
We fix some notations first. Let \[\label{eqn:fibxs} p: X \longrightarrow S\] be a Weierstraß fibration in the sense of. The fibration has a section \(\sigma: S \to X\) whose image lies in the smooth locus of the morphism \(p\). We write \(\Theta = \sigma(S)\) and \(h = p|_{\Theta}: \Theta \to S\). Then \(h\) is clearly an isomorphism. In the case of our interest, we further assume that both \(X\) and \(S\) are smooth. We label the relevant morphisms as in the following fiber diagram \[X \times_S X \ar[r]^-{\pi_2} \ar[d]_-{\pi_1} \ar[rd]^-{\pi} & X \ar[d]^p \\ X \ar[r]_p & S. }\] We define an integral functor whose kernel is given by the sheaf \[\label{eqn:FM-kernel} \mathcal{P} = \mathcal{I}_{\Delta} \otimes \pi_1^*\mathcal{O}_X(\Theta) \otimes \pi_2^*\mathcal{O}_X(\Theta) \otimes \pi^*\omega^{-1}\] where \(\mathcal{I}_{\Delta}\) is the ideal sheaf of the image \(\Delta\) of the diagonal morphism \(\delta: X \hookrightarrow X \times_S X\), and \[\omega = R^1p_*\mathcal{O}_X\] as introduced in. Then we have the following result:
The Fourier-Mukai transform \(\Phi\) defined in the above theorem will be one of the main functors that we will consider. Now we look at the other functor that will play a key role.
## The dualizing functor
Let \(Z\) be a Gorenstein variety and \(\mathcal{E} \in D^b(Z)\). We define \[\Delta_Z(\mathcal{E}) := \shom_{\mathcal{O}_Z}(\mathcal{E}, \mathcal{O}_Z).\] It is familiar that when \(Z\) is smooth, we have \[\label{eqn:dual-bounded} \Delta_Z(\mathcal{E}) \in D^b(Z)\] and \[\label{eqn:dual-dual} (\Delta_Z \circ \Delta_Z)(\mathcal{E})=\mathcal{E}.\] In fact, both relations still hold when \(Z\) is Gorenstein; see for example. We will use the dualizing functor on \(X\) and \(X \times_S X\). Notice that \(X \times_S X\) is in general singular, but still Gorenstein. Indeed, the projection morphism \(\pi_1: X \times_S X \to X\) is a base change of the morphism \(p: X \to S\). Since \(p: X \to S\) is flat with Gorenstein fibers, the same is true for \(\pi_1: X \times_S X \to X\), hence they are both Gorenstein morphisms; see. It follows that \(X \times_S X\) is Gorenstein by transitivity. Hence both [\[eqn:dual-bounded\]](#eqn:dual-bounded){reference-type="eqref" reference="eqn:dual-bounded"} and [\[eqn:dual-dual\]](#eqn:dual-dual){reference-type="eqref" reference="eqn:dual-dual"} hold for \(X\) and \(X \times_S X\).
## The commutativity of functors
Now we discuss the commutativity of the Fourier-Mukai functor and the dualizing functors. Recall that for the Weierstraß fibration [\[eqn:fibxs\]](#eqn:fibxs){reference-type="eqref" reference="eqn:fibxs"}, an \(S\)-involution is a morphism \(\iota: X \to X\) satisfying \(\iota \circ \iota = \id_X\) and \(p \circ \iota = p\). The key result concerning the commutativity of functors is the following:
This is a very convenient tool for our discussion in next section. As pointed out in, this result generalize the classical result of Mukai in. Before discussing its applications we make some remarks concerning \(L\) and \(\iota\) in the statement of the theorem.
# Spectral sequences and applications {#sec:specseq}
In this section we will obtain some consequences of Theorem [\[thm:commutativity\]](#thm:commutativity){reference-type="ref" reference="thm:commutativity"} for the Weierstraß fibration [\[eqn:fibxs\]](#eqn:fibxs){reference-type="eqref" reference="eqn:fibxs"}. More precisely, we will apply spectral sequences on both sides of [\[eqn:2comp\]](#eqn:2comp){reference-type="eqref" reference="eqn:2comp"} and obtain some identities by comparing their cohomology. These identities will be used in later sections to determine the behavior of line bundles under Fourier-Mukai transforms. For simplicity, given any \(E \in \coh(X)\), the \(i\)-th cohomology sheaf of \(\Phi(E)\) is denoted by \(\Phi^iE\). We also write \(n\) for \(\dim X\). All functors in the spectral sequences are underived.
## The duality spectral sequences
The composition of derived functors gives spectral sequences. The identity [\[eqn:2comp\]](#eqn:2comp){reference-type="eqref" reference="eqn:2comp"} in Theorem [\[thm:commutativity\]](#thm:commutativity){reference-type="ref" reference="thm:commutativity"} shows that two spectral sequences converge to the same limit. More precisely, for any \(E \in \coh(X)\), the \(E_2\)-page of the spectral sequence corresponding to the composition functor \(\Delta_X \circ \Phi\) is given by \[E_2^{p,q} = \sext^q_{\mathcal{O}_X}(\Phi^{-p}E, \mathcal{O}_X) \Longrightarrow E_{\infty}.\] Notice that the non-trivial region for \((p,q)\) in the \(E_2\)-page is bounded by \(-1 \leq p \leq 0\) and \(0 \leq q \leq n\). Similarly, the \(E_2\)-page of the spectral sequence corresponding to the composition functor \(\Phi \circ \Delta_X\) is given by \[E_2'^{p,q} = \Phi^q\sext^p_{\mathcal{O}_X}(E, \mathcal{O}_X) \Longrightarrow E'_{\infty},\] with the non-trivial region for \((p,q)\) in the \(E_2\)-page bounded by \(0 \leq p \leq n\) and \(0 \leq q \leq 1\). Taking the shift functor and the involution \(\iota\) into consideration, we get the following result
These two spectral sequences provide us lots of information between the sheaf \(E\) and its dual. More precisely, after stabilization we can compare the anti-diagonals which gives the same term in the limit. In some cases, we know the arrows among the terms \(E_{2L}^{pq}\) are also trivial, then we can simply extract information from the \(E_2\)-pages. Some examples of such analysis are given below.
## Applications of the duality spectral sequence
As an application of Proposition [\[prop:comm\]](#prop:comm){reference-type="ref" reference="prop:comm"}, we prove the following two interesting statements about sheaves with WIT properties. We point out that in these examples both spectral sequences degenerate at \(E_2\)-page, because we will see that the first spectral sequence will have only one column of non-trivial terms on the \(E_2\)-page. Before we state the first proposition, we collect some facts and notations. Recall that for any \(\mathcal{E} \in D^b(X)\), the dimensions of the supports of \(\mathcal{E}\) and \(\Phi(\mathcal{E})\) differ at most by \(1\); i.e. \(\dim \mathcal{E}-1 \leq \dim \Phi(\mathcal{E}) \leq \dim \mathcal{E} +1\). This follows from. Moreover, for any \(E \in \coh(X)\), it is proven in that \(\Phi^iE=0\) unless \(i=0\) or \(1\). If we assume further that the support of \(E\) is of codimension \(c\), then we have that \(\sext^i(E, \mathcal{O}_X)=0\) for all \(i<c\) by. We define the dual sheaf of \(E\) by \(E^D=\sext^c(E, \mathcal{O}_X)\). We will discuss the consequences of the dualizing spectral sequence for each possibility of the difference in the dimensions assuming \(E\) is a \(\Phi\)-WIT\(_0\) or \(\Phi\)-WIT\(_1\) sheaf on \(X\). Recall that a sheaf \(E \in \coh(X)\) is said to be \(\Phi\)-WIT\(_i\) if \(\Phi(E)[i] \in \coh(X)\).
Now we turn to sheaves with \(\Phi\)-WIT\(_1\) property. We can similarly prove the following result.
Finally, we prove a very simple result along the same line, but without assuming any WIT property of \(E\). We can think of it as a result which is slightly stronger than Propositions [\[prop:wit0\]](#prop:wit0){reference-type="ref" reference="prop:wit0"} and [\[prop:wit1\]](#prop:wit1){reference-type="ref" reference="prop:wit1"} in the special case of \(\dim E = \dim X = n\).
The above results will be used to study the behavior of sheaves under Fourier-Mukai transforms in the next section.
# Transforms of line bundles {#sec4}
We now use the duality spectral sequences from Section [3](#sec:specseq){reference-type="ref" reference="sec:specseq"} to study line bundles under Fourier-Mukai transforms on elliptic fibrations.
## The basic setting {#subsec:setting}
In this section, we compute \(\ch_0, \ch_1\) of the transforms of certain line bundles, so that we can study the slope stability of these transforms later on. We will now make the following assumptions on our fibration \(p: X \to S\):
- The total space \(X\) in our Weierstraß fibration \(p: X \to S\) is a \(K\)-trivial threefold.
- The canonical class \(K_S\) of the base \(S\) of the fibration \(p\) is *numerically trivial*. This allows \(S\) to be a K3 surface or an Enriques surface.
- The cohomology ring of \(X\) is of the form \[H^{2i}(X,\mathbb{Q}) = \Theta p^\ast H^{2i-2}(S,\mathbb{Q}) \oplus p^\ast H^{2i}(S,\mathbb{Q}),\] an assumption often made in the study of elliptic threefolds as in.
Note that applying adjunction to the closed immersion \(\Theta \hookrightarrow X\) gives \[\label{eqn:Sigma-sq} \Theta^2 = \Theta\cdot p^\ast K_S.\] For any integer \(m\), we will write \[L_m = \mathcal{O}_X (m\Theta).\] That \(K_S\) is numerically trivial in assumption (ii) implies \(\Theta^2=0\) and \(\Theta^3=0\) by [\[eqn:Sigma-sq\]](#eqn:Sigma-sq){reference-type="eqref" reference="eqn:Sigma-sq"}. Hence by the formulas in, we obtain \(\ch (L_m) = 1 + m\Theta\) and \[\begin{aligned}
\ch_0 (\Phi L_m) &= m, \\ \ch_1 (\Phi L_m) &=-\Theta-\tfrac{m}{2}c \end{aligned}\] where \(c =-p^\ast K_S\). We can now compute the slope of the transform of \(L_m\). Let us fix an ample class \(\omega\) on \(X\). By assumption (iii) above, we know \(\omega\) must be of the form \[\label{def:omega} \omega = t\Theta + s p^\ast H_S\] for some ample class \(H_S\) on \(S\), and some real numbers \(t, s >0\). In this notation \[\omega^2 = t^2 \Theta^2 + 2ts \Theta p^\ast H_S + s^2 p^\ast H_S^2.\] The slope of \(\Phi L_m\) with respect to the polarisation \(\omega\) is then \[\begin{aligned}
\mu_\omega (\Phi L_m) &= \frac{\ch_1(\Phi L_m)\omega^2}{\ch_0 (\Phi L_m)} \notag\\ &= \frac{(-\Theta-\tfrac{m}{2}c)(t^2\Theta^2+2ts\Theta p^\ast H_S + s^2 p^\ast H_S^2)}{m} \\ &= \frac{-s^2 \Theta p^\ast H_S^2}{m} \\ &=-\frac{s^2}{m}H_S^2. \label{eqn:L-aslope} \end{aligned}\]
## Transforms of reflexive sheaves
As above, we write \(L_m = \mathcal{O}_X (m\Theta)\).
## *K*-trivial elliptic threefolds over numerically *K*-trivial surfaces | {'timestamp': '2017-06-08T02:01:12', 'yymm': '1706', 'arxiv_id': '1706.01923', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01923'} |
# Introduction {#sec_introduction}
Generalized linear mixed models (GLMMs) are generalized linear models with random terms in the linear predictor. The random effects in the GLMM can accommodate for overdispersion often present in non-Gaussian data, and dependence among correlated observations arising from longitudinal or repeated measures studies. GLMM is one of the most frequently used statistical models. Here, we consider a popular Bayesian GLMM for binary data, namely, the probit linear mixed model. Let \((Y_1, Y_2, \dots, Y_n)\) denote the vector of Bernoulli random variables. Let \(\bm{x}_i\) and \(\bm{z}_i\) be the \(p \times 1\) and \(q \times 1\) known covariates and random effect design vectors respectively associated with the \(i\)th observation for \(i=1,\dots,n\). Let \(\bm{\beta} \in \mathbb{R}^p\) be the unknown vector of regression coefficients and \(\bm{u} \in \mathbb{R}^q\) be the random effects vector. A GLMM can be built with a link function that connects the expectation of \(Y_i\) with \(\bm{x}_i\) and \(\bm{z}_i\). One of the very popular link functions is the probit link function, \(\Phi^{-1}\), resulting in \[\label{eq_link} P(Y_i =1) = \Phi(\bm{x}_i^T\bm{\beta} + \bm{z}_i^T\bm{u}),\] where \(\Phi(\cdot)\) is the cumulative distribution function of the standard normal random variable. Assume that we have \(r\) random effects with \(\bm{u} = (\bm{u}_1^T,\dots, \bm{u}_r^T)^T\), where \(\bm{u}_j\) is a \(q_j\times1\) vector with \(q_j >0\), \(q_1+\cdots + q_r = q\), and \(\bm{u}_j \stackrel{\text{ind}}{\sim} N(0, \bm{I}_{q_j}1/\tau_j)\), where \(\tau_j\in \mathbb{R}_+ \equiv (0,\infty)\) is the precision parameter associated with \(\bm{u}_j\) for \(j=1,\dots,r\). Let \(\bm{\tau}= (\tau_1, \dots, \tau_r)\), thus, the data model for the probit GLMM is \[\begin{aligned}
\label{eq:data} Y_{i} | \bm{\beta}, \bm{u}, \bm{\tau}& \overset{\text{ind}}\sim & \text{Bern}(\alpha_{i}) \text{ for } i=1,\dots,n \;\text{ with} \nonumber\\ \alpha_{i} & = & \Phi(\bm{x}_{i}^{T}\bm{\beta}+\bm{z}_{i}^{T}\bm{u}) \text{ for }i=1,\dots,n, \\ \bm{u}_{j}| \bm{\beta}, \bm{\tau} & \overset{\text{ind}}\sim & N\Big(0,\frac{1}{\tau_{j}}\bm{I}_{q_{j}}\Big), \, j=1,\dots,r.\nonumber \end{aligned}\] Let \(\bm{y} = (y_1, y_2, \dots, y_n)^T\) be the observed Bernoulli response variables. Note that, the likelihood function for \((\bm{\beta}, \bm{\tau})\) is \[\begin{aligned}
\label{eq:lik} L(\bm{\beta}, \bm{\tau} | \bm{y}) &=& \int_{\mathbb{R}^{q}} \prod_{i=1}^n\left[\Phi(\bm{x}_i^T\bm{\beta} + \bm{z}_i^T\bm{u})\right]^{y_i}\left[1-\Phi(\bm{x}_i^T\bm{\beta} +\bm{z}_i^T\bm{u})\right]^{1-y_i} \\ & \times & \phi_q(\bm{u}; \bm{0}, \bm{D}(\bm{\tau})^{-1})d\bm{u}, \nonumber \end{aligned}\] which is not available in closed form. Here, \(\phi_q(s; a, B)\) denotes the probability density function of the \(q-\)dimensional normal distribution with mean vector \(a\), covariance matrix \(B\) and evaluated at \(s\), and \(\bm{D}(\bm{\tau}) = \oplus_{j=1}^r \tau_j \bm{I}_{q_j}\). In Bayesian framework, one needs to specify the prior distributions of \(\bm{\beta}\) and \(\bm{\tau}\). Assume \(\bm{\beta}\) and \(\bm{\tau}\) are apriori independent. Let \(\pi(\bm{\beta})\) and \(\pi(\bm{\tau})\) be the prior densities of \(\bm{\beta}\) and \(\bm{\tau}\) respectively. Thus, the joint posterior density of \((\bm{\beta},\bm{\tau})\) is \[\label{eq:postbetphi} \pi(\bm{\beta},\bm{\tau}|\bm{y}) = \frac{1}{c(\bm{y})} L(\bm{\beta}, \bm{\tau}| \bm{y})\pi(\bm{\beta})\pi({\bm{\tau}}),\] where \[c(\bm{y}) = \int_{\mathbb{R}_+^{r}} \int_{\mathbb{R}^{p}} L(\bm{\beta}, \bm{\tau}| \bm{y}) \pi(\bm{\beta}) \pi(\bm{\tau})d\bm{\beta} d\bm{\tau},\] is the marginal density of \(\bm{y}\). Since the likelihood function \(L(\bm{\beta}, \bm{\tau}| \bm{y})\) is not available in closed form, the posterior density is intractable for any choice of the prior distributions of \(\bm{\beta}\) and \(\bm{\tau}\). In this article, we consider an improper flat prior for \(\bm{\beta}\), that is, \(\pi(\bm{\beta})\propto 1\) and \(\tau_j\)'s, \(j=1,\dots, r\), are apriori independent with \[\label{eq:improper_prior} \pi(\tau_j )\propto e^{-b_j\tau_j}\tau_j^{a_j-1},\] which can be proper or improper. In section [2](#sec_propriety){reference-type="ref" reference="sec_propriety"}, we discuss conditions under which the posterior density [\[eq:postbetphi\]](#eq:postbetphi){reference-type="eqref" reference="eq:postbetphi"} is proper, that is \(c(\bm{y}) < \infty\). Generally, Markov chain Monte Carlo (MCMC) algorithms are used for exploring the posterior density [\[eq:postbetphi\]](#eq:postbetphi){reference-type="eqref" reference="eq:postbetphi"}. Even in the absence of random effects, for the probit regression model, the posterior distribution of \(\bm{\beta}\) is difficult to sample from. MCMC algorithm for sampling from the posterior distribution associated with the probit regression model is the most widely used data augmentation (DA) algorithm. The DA technique used in can also be applied to the probit linear mixed model. Following, let \(v_i \in \mathbb{R}\) be the continuous latent variable corresponding to the \(i\)th binary observation \(Y_i\), such that \(Y_i = I(v_i >0)\), where \(v_i | \bm{\beta}, \bm{u}, \bm{\tau}\overset{\text{ind}} \sim N(\bm{x}_i^{T} \bm{\beta} + \bm{z}_{i}^{T}\bm{u}\), 1) for \(i =1,\dots, n\). Then \[\label{eq:vy} P(Y_i=1) = P(v_i >0) =\Phi(\bm{x}_i^{T} \bm{\beta} + \bm{z}_{i}^{T}\bm{u}),\] that is, \(Y_{i} | \bm{\beta}, \bm{u}, \bm{\tau} \overset{\text{ind}}\sim \text{Bern}(\alpha_{i})\) as in [\[eq:data\]](#eq:data){reference-type="eqref" reference="eq:data"}. Note that \(\bm{v}| \bm{\beta}, \bm{u}, \bm{\tau} \sim N(\bm{X}\bm{\beta} + \bm{Z}\bm{u},\bm{I}_n)\), where \(\bm{v} = (v_1,\dots,v_n)^T\), \(\bm{X}_{n\times p} =(\bm{x}_1,\dots, \bm{x}_{n})^T\) and \(\bm{Z}_{n\times q} = (\bm{z}_1,\dots, \bm{z}_n)^{T}\). Using the latent variables \(\bm{v}\), we can introduce a joint density \(\pi(\bm{\beta}, \bm{u},\bm{v},\bm{\tau}|\bm{y})\) (see section [3](#sec_twoblock){reference-type="ref" reference="sec_twoblock"} for details) such that \[\label{eq:margjt} \int_{\mathbb{R}^{q}} \int_{\mathbb{R}^{n}} \pi(\bm{\beta}, \bm{u},\bm{v},\bm{\tau}|\bm{y}) d\bm{v}d\bm{u} = \pi(\bm{\beta},\bm{\tau}|\bm{y}),\] where \(\pi(\bm{\beta},\bm{\tau}|\bm{y})\) is the posterior density defined in [\[eq:postbetphi\]](#eq:postbetphi){reference-type="eqref" reference="eq:postbetphi"}. If all the full conditionals of the joint density \(\pi(\bm{\beta}, \bm{u},\bm{v},\bm{\tau}|\bm{y})\) are easy to sample from, then a Gibbs sampler can be run and it can be used to make inferences on the posterior density [\[eq:postbetphi\]](#eq:postbetphi){reference-type="eqref" reference="eq:postbetphi"}. Indeed this full Gibbs sampler is traditionally used in the analysis of Bayesian probit linear mixed models. In this article, instead of using full conditional distributions, we construct a two-block Gibbs sampler with \(\bm{\eta} \equiv (\bm{\beta}^T,\bm{u}^T)^T\) as one block and \((\bm{v}^T, \bm{\tau}^T)^T\) as the other block---which is our first contribution. In general, block Gibbs samplers are known to be better than the Gibbs samplers based on full conditional distributions in terms of having smaller operator norm. The above mentioned block Gibbs sampler has an everywhere strictly positive Markov transition density, implying that the underlying Markov chain is Harris ergodic [@asmu:glyn:2011; @meyn1993markov]. Thus, the time average estimators based on the block Gibbs sampler can be used to consistently estimate the (posterior) means with respect to the joint density \(\pi(\bm{\beta}, \bm{u},\bm{v},\bm{\tau}|\bm{y})\). In practice, it is crucial to know whether the Monte Carlo errors associated with these estimates are sufficiently small. However, in order to provide valid standard errors, we need to establish a central limit theorem (CLT) for the time average estimators. Unlike for the ordinary Monte Carlo methods based on iid samples, mere existence of the finite second moment does not guarantee a CLT for MCMC estimators. One standard method of establishing a CLT for MCMC estimators is to prove that the underlying Markov chain is *geometrically ergodic*. Geometric ergodicity is also needed for consistently estimating the asymptotic variance in the Markov chain CLT. and proved geometric ergodicity of DA algorithm for the Bayesian probit regression model under improper and proper priors on the regression coefficients. For linear models, and analyzed the Gibbs sampler for one-way random effects models under proper priors and improper priors respectively. analyzed the block Gibbs sampler for Bayesian linear mixed models under the assumption \(\bm{X}^T\bm{Z} = \bm{0}\). and established geometric rate of convergence of the Gibbs samplers for Bayesian linear mixed models under improper and proper priors without the assumption of \(\bm{X}^T\bm{Z} = \bm{0}\). Our second contribution, in this paper, is establishing geometric convergence rates for the block Gibbs sampler for Bayesian probit linear mixed models under improper priors. DA algorithms are known to suffer from slow convergence. proposed the *parameter expansion for data augmentation* (PX-DA) algorithm, which can converge faster than the DA algorithm without much extra computational effort [@vand:meng:2001; @roy2014efficient]. proved that the Haar PX-DA algorithm, that is based on a Haar measure, is better than any other PX-DA algorithm and the original DA algorithm in both the efficiency ordering and the operator norm ordering. For the probit regression model,, through an example, showed that the Haar PX-DA algorithm can lead to huge gains in efficiency over the DA algorithm of. Our third contribution is to construct a Haar PX-DA algorithm improving the block Gibbs sampler mentioned before. Since geometric ergodicity of the Haar PX-DA algorithm follows from geometric ergodicity of the DA algorithm, we have CLTs for the Haar PX-DA algorithm based estimators as well. The article is organized as follows. In section [2](#sec_propriety){reference-type="ref" reference="sec_propriety"}, we establish conditions for propriety of the posterior distribution [\[eq:postbetphi\]](#eq:postbetphi){reference-type="eqref" reference="eq:postbetphi"} under improper priors, when \(\bm{X}\) and \(\bm{Z}\) take commonly observed forms. The results in section [2](#sec_propriety){reference-type="ref" reference="sec_propriety"} hold for a general link function, not necessarily the probit link. In section [3](#sec_twoblock){reference-type="ref" reference="sec_twoblock"}, we construct the two-block Gibbs sampler for the Bayesian probit linear mixed model under improper priors. In section [4](#sec_convergence){reference-type="ref" reference="sec_convergence"}, we prove geometric ergodicity of the underlying Markov chain. In section [5](#sec_px-da){reference-type="ref" reference="sec_px-da"}, we present a corresponding Haar PX-DA algorithm. Section [6](#sec_conclusion){reference-type="ref" reference="sec_conclusion"} contains some conclusions and discussions. Finally, the proofs of posterior propriety and geometric convergence of the Gibbs sampler appear in the appendices.
# Propriety of posterior distributions {#sec_propriety}
In this section, we discuss conditions under which the posterior density [\[eq:postbetphi\]](#eq:postbetphi){reference-type="eqref" reference="eq:postbetphi"} is proper. The results in this section hold for GLMMs with a general link function. Let \(F(\cdot)\) be a cumulative distribution function, and consider the link function \(F^{-1}(\cdot)\). Thus instead of the probit linear mixed model in [\[eq_link\]](#eq_link){reference-type="eqref" reference="eq_link"}, in this section we consider a GLMM with \[\label{eq_link_general} P(Y_i =1) = F(\bm{x}_i^T\bm{\beta} + \bm{z}_i^T\bm{u}).\] Posterior propriety for Bayesian GLMMs under improper priors has been discussed in. We will first describe conditions. Then we will show, through examples, that these conditions often do not hold in practice. Finally, our conditions for posterior propriety will be presented. Let \(c_i = 1\) if \(y_i = 0\) and \(c_i =-1\) if \(y_i = 1\) for \(i=1,\dots, n\). Suppose \(\bm{W}_{n\times(p+q)}^*\) is a matrix whose \(i\)th row is \(c_i(\bm{x}_i^T,\bm{z}_i^T)\). In the special case when \(b_j=0\), that is, when \(\tau_j\) has the power prior \(\pi(\tau_j) \propto \tau_j^{a_j-1}\) for \(j=1,\dots,r\), a straightforward extension of Theorem 4.2 shows that the corresponding posterior distribution is proper if the following conditions hold:
1. \(\bm{W} = (\bm{X}, \bm{Z})\) is a full rank matrix;
2. There exists an \(n\times 1\) positive vector \(\bm{e}>0\) such that \(\bm{e}^T\bm{W}^* = 0\);
3. \(2a_j+ q_j >0\) for \(j=1,2,\dots,r\);
4. \(a_j < 0\) for \(j=1,\dots, r\);
5. \(E \vert \delta \vert^{p-2\sum_{j=1}^ra_j} <\infty\), where \(\delta \sim F\).
provided a simple method for checking the condition A2 using publicly available softwares. The condition A1 assumes that \(\bm{W}\) is a full rank matrix. Unfortunately, when \(\bm{Z}\) is a design matrix with elements 1's and 0's, which is pretty common in practice, this assumption may not hold. For example, we consider the following important generalized two-way random effects model \[\label{eq_link_example} F^{-1}(P(Y_{ij} = 1)) = \beta + \alpha_i + \gamma_j,\] for \(i = 1,2, \dots, n_1\), \(j = 1,2,\dots, n_2\). Here, the \(\alpha_i\)'s are i.i.d \(N(0,1/\tau_1)\), and the \(\gamma_j\)'s are i.i.d \(N(0,1/\tau_2)\). There are total \(n = n_1\times n_2\) observations and we order them as \(\bm{Y} = (Y_{11},\dots, Y_{1n_2}, \dots, Y_{n_11}, \dots, Y_{n_1n_2})\). In this example, \(p=1\), and \(\bm{X} = \bm{1}_n\) is an \(n\times 1\) column vector of ones. Also, there are \(r=2\) random effects with \(q_1 = n_1\), \(q_2 = n_2\), \(q = q_1 + q_2 = n_1 + n_2\), and \(\bm{Z} = (\bm{Z}_1, \bm{Z}_2 )\), where \(\bm{Z}_1 = \bm{I}_{n_1} \otimes \bm{1}_{n_2}\) and \(\bm{Z}_2 = \bm{1}_{n_1}\otimes \bm{I}_{n_2}\) with \(\otimes\) denoting the Kronecker product. It can be checked that the rank of \(\bm{W} = (\bm{X},\bm{Z})\) is \(n_1 + n_2-1\). Thus \(\bm{W}\) is not a full rank matrix. We now provide Theorem [\[them3\]](#them3){reference-type="ref" reference="them3"} showing the posterior propriety without the assumption A1. We also consider the more general prior \(\pi(\tau_j)\) given in [\[eq:improper_prior\]](#eq:improper_prior){reference-type="eqref" reference="eq:improper_prior"}, that is, \(b_j\) may not be zero. We use certain transformations of the regression parameters \(\bm{\beta}\) and random effects \(\bm{u}\) to circumvent the problem with non-full rank matrix \(\bm{W}\). Assume that the first column of \(\bm{X}\) is a vector of 1's corresponding to an intercept term \(\beta_0\) in \(\bm{\beta} = (\beta_0, \beta_1,\dots, \beta_{p-1})^T\). Let \(\bm{Z} = (\bm{Z}_1,\dots,\bm{Z}_r)\), where \(\bm{Z}_j\) is an \(n\times q_j\) matrix such that the \((ik)\)th element is 1 if the observation \(i\) is observed at the \(k\)th level of the random effect \(\bm{u}_{j}=\left(u_{j1},\dots,u_{jq_{j}}\right)^{T}\), 0 otherwise, for \(i=1, \dots, n\), \(k=1,\dots, q_j\) and \(j=1,\dots, r\). Consider the following transformations, \[\begin{aligned}
\label{eq_trans1} \mu_{0} &=&\beta_{0}+\sum_{j=1}^{r}u_{j1},\\ \label{eq_trans2} d_{jk} &=&u_{j,k+1}-u_{j1}, \text{for } k=1,\dots,q_{j}-1,j=1,\dots,r. \end{aligned}\] Thus \(\mu_0\) is the sum of the intercept term and the first level effect of all \(r\) random effects. Also the (transformed) random effects \(d_{jk}\)'s denote the differences of the random effect compared to the first level effect. Let \(\tilde{\bm{\eta}}=\left(\mu_{0},\beta_{1,},\dots,\beta_{p-1},d_{11},\dots,d_{1,q_{1}-1},\dots d_{r1},\dots,d_{r,q_{r}-1}\right)^{T}\). Define \(\tilde{\bm{Z}} = (\tilde{\bm{Z}}_1,\dots, \tilde{\bm{Z}}_r)\), where the \(n\times (q_j-1)\) matrix \(\tilde{\bm{Z}}_j\) is \(\bm{Z}_j\) without its first column. Thus, the vector \(\bm{W}\bm{\eta}\) is the same as the vector \(\tilde{\bm{W}}\tilde{\bm{\eta}}\), where \(\tilde{\bm{W}}=\left(\bm{X},\tilde{\bm{Z}}\right)\) with \(i\)th row \(\tilde{\bm{w}}_i^T\). Let \(\tilde{\bm{W}}^*\) be a matrix whose \(i\)th row is \(\tilde{\bm{w}}_i^{*T} =c_i\tilde{\bm{w}}_i^T = c_i(\bm{x}_i^T, \tilde{\bm{z}}_i^T)\), where \(\tilde{\bm{z}}_{i}^T\) is the \(i\)th row of \(\tilde{\bm{Z}}\). For the example [\[eq_link_example\]](#eq_link_example){reference-type="eqref" reference="eq_link_example"}, the transformed parameters \(\mu_0\) and \(d_{jk}\)'s become \[\begin{aligned}
\mu_0 & =& \beta + \alpha_1 + \gamma_1,\\ d_{1k} &=& \alpha_{k+1}-\alpha_1 \text{ for } k=1,\dots,n_1-1,\\ d_{2k} &=& \gamma_{k+1}-\gamma_1 \text{ for } k=1,\dots,n_2-1. \end{aligned}\] Thus in this example, we have \(\tilde{\bm{\eta}} = (\mu_0, d_{11},\dots, d_{1,n_1-1}, d_{21},\dots, d_{2,n_2-1})^T\). Also note that \(\tilde{\bm{W}}\) is a full rank matrix in this example, although \(\bm{W}\) is not.
A proof of Theorem [\[them3\]](#them3){reference-type="ref" reference="them3"} is given in Appendix [\[proof_thm3\]](#proof_thm3){reference-type="ref" reference="proof_thm3"}.
# A two-block Gibbs sampler {#sec_twoblock}
We begin with deriving the joint density \(\pi(\bm{\beta}, \bm{u},\bm{v},\bm{\tau}|\bm{y})\) mentioned in the introduction. Define the joint posterior density (up to a normalizing constant) of \(\bm{\beta}, \bm{u}, \bm{v}, \bm{\tau}\), if it exists, as \[\begin{aligned}
\label{eq_joint_improper} \pi(\bm{\beta},\bm{u}, \bm{v}, \bm{\tau}|\bm{y}) & \propto \prod_{i=1}^n \exp\left\{-\frac{1}{2}\left( v_i-\bm{x}_i^T\bm{\beta}-\bm{z}_i^T\bm{u} \right)^2 \right\} \nonumber\\ & \times \prod_{i=1}^n \left[1_{\text{(0,\ensuremath{\infty})}}\left(v_{i}\right)\right]^{y_{i}}\left[1_{\left(-\infty,0\right]}\left(v_{i}\right)\right]^{1-y_{i}} \nonumber \\ & \times \prod_{j=1}^r \tau_j^{\frac{q_j}{2}+a_j-1}\exp\left\{-\tau_j\left( b_j + \frac{\bm{u}_j^T\bm{u_j}}{2}\right)\right\}. \end{aligned}\] From [\[eq:lik\]](#eq:lik){reference-type="eqref" reference="eq:lik"} and [\[eq:vy\]](#eq:vy){reference-type="eqref" reference="eq:vy"} it follows that [\[eq:margjt\]](#eq:margjt){reference-type="eqref" reference="eq:margjt"} holds. In section [2](#sec_propriety){reference-type="ref" reference="sec_propriety"}, we discussed conditions under which the posterior density \(\pi(\bm{\beta},\bm{\tau}\vert \bm{y})\) given in [\[eq:postbetphi\]](#eq:postbetphi){reference-type="eqref" reference="eq:postbetphi"} and hence the joint posterior density [\[eq_joint_improper\]](#eq_joint_improper){reference-type="eqref" reference="eq_joint_improper"} is proper. Note that, these posterior densities are proper if and only if \(c(\bm{y}) < \infty\). Standard calculations show that the conditional density of \(\bm{\eta}\) is \[\label{eq_posterior1} \pi(\bm{\eta}|\bm{v},\bm{\tau},\bm{y}) \propto \exp\left[-\frac{1}{2}\left(\bm{v}-\bm{W}\bm{\eta}\right)^{T}\left(\bm{v}-\bm{W}\bm{\eta}\right)\right] \cdot\exp\left[-\frac{1}{2} \bm{u}^{T} \bm{D}(\bm{\tau}) \bm{u} \right],\] Thus, \[\label{eq:eta} \bm{\eta}|\bm{v}, \bm{\tau},\bm{y}\sim N_{p+q}\left(\bm{\bm{\Sigma}}^{-1}\bm{W}^T\bm{v},\bm{\Sigma}^{-1}\right),\] where \[\label{eq:improper_eta_sigma} \bm{\Sigma}=\left(\begin{array}{cc} \bm{X}^{T}\bm{X} & \bm{X}^{T}\bm{Z}\\ \bm{Z}^{T}\bm{X} & \bm{Z}^{T}\bm{Z}+\bm{D}(\bm{\tau}) \end{array}\right).\] Similarly, the conditional density of \((\bm{v}, \bm{\tau})\) is \[\begin{aligned}
\pi(\bm{v},\bm{\tau}|\bm{\eta},\bm{y}) & \propto & \prod_{i=1}^{n} \phi \left(v_i-\bm{w}_i^T\bm{\eta};0,1\right)\left[1_{\text{(0,\ensuremath{\infty})}}\left(v_{i}\right)\right]^{y_{i}}\left[1_{\left(-\infty,0\right]}\left(v_{i}\right)\right]^{1-y_{i}}\\ & & \times\prod_{j=1}^{r}\tau_{j}^{\frac{q_{j}}{2} + a_j-1}\exp\left[-\tau_{j} \left(b_j + \frac{1}{2}\bm{u}_{j}^{T}\bm{u}_{j}\right) \right], \end{aligned}\] where \(\bm{w}_i^T\) is the \(i\)th row of \(\bm{W}\) for \(i=1,\dots,n\). Thus, conditional on \((\bm{\eta}, \bm{y})\), \(v_i\), \(i = 1, \dots, n\) and \(\bm{\tau}\) are independent. We have \[\label{eq:v} v_i| \bm{\eta}, \bm{y} \overset{\text{ind}}\sim \text{TN}(\bm{w}_i^T\bm{\eta},1,y_i), \, i=1,\dots, n,\] where \(\text{TN}(\mu, \sigma^2, \omega)\) denotes the distribution of the normal random variable with mean \(\mu\) and variance \(\sigma^2\), that is truncated to have only positive values if \(\omega = 1\), and only nonpositive values if \(\omega = 0\). Also conditional on \(\bm{\eta}, \bm{y}\), \(\tau_j\)'s are independent with \(\tau_j \sim \text{Gamma}\left(a_j+q_j/2, b_j+\bm{u}_j^T\bm{u}_j/2\right)\) for \(j=1,\dots,r\). Thus, one single iteration of the block Gibbs sampler \(\{\bm{\eta}^{(m)}, \bm{v}^{(m)}, \bm{\tau}^{(m)}\}_{m=0}^{\infty}\) has the following two steps:
A proof of Theorem [\[them2\]](#them2){reference-type="ref" reference="them2"} is given in the Appendix [\[proof_thm2\]](#proof_thm2){reference-type="ref" reference="proof_thm2"}. Theorem [\[them4\]](#them4){reference-type="ref" reference="them4"} shows geometric convergence of the Markov chain underlying the Gibbs sampler given in Algorithm [\[algorithm2\]](#algorithm2){reference-type="ref" reference="algorithm2"} without the assumption A1.
A proof of Theorem [\[them4\]](#them4){reference-type="ref" reference="them4"} is given in Appendix [\[proof_thm4\]](#proof_thm4){reference-type="ref" reference="proof_thm4"}.
# A Haar PX-DA algorithm {#sec_px-da}
As mentioned in section [1](#sec_introduction){reference-type="ref" reference="sec_introduction"}, DA algorithms often suffer from slow convergence and high autocorrelations. proposed parameter expansion for data augmentation (PX-DA) algorithms for speeding up the convergence of DA algorithms. compared the performance of PX-DA algorithms based on a Haar measure (called Haar PX-DA algorithms) with PX-DA algorithms based on a probability measure and DA algorithms. In particular, they showed that, under some mild conditions, the Haar PX-DA algorithms are better than the general PX-DA algorithms and the DA algorithms in both the efficiency ordering and the operator norm ordering. As shown in, compared to the DA algorithm, in PX-DA, an extra step is added (sandwiched) between the two steps of the original DA algorithm. In order to construct this extra step, we derive the marginal density \[\begin{aligned}
\label{eq:margvtau1} \pi\left(\bm{v}, \bm{\tau}|\bm{y}\right) & = & \int_{\mathbb{R}^{p+q}}\pi(\bm{\eta},\bm{v},\bm{\tau}|\bm{y})d\bm{\eta}\\ & \propto & \prod_{i=1}^{n}\left[1_{\text{(0,\ensuremath{\infty})}}\left(v_{i}\right)\right]^{y_{i}}\left[1_{\left(-\infty,0\right]}\left(v_{i}\right)\right]^{1-y_{i}} \prod_{j=1}^{r}\tau_{j}^{\frac{q_{j}}{2}+a_j} e^{-b_j\tau_j}\nonumber\\ & & \cdot \vert \bm{\Sigma}\vert^{-1/2}\exp\left\{-\frac{1}{2}\bm{v}^{T}M_1\bm{v}\right\}, \nonumber \end{aligned}\] where \(M_1 = \bm{I}-\bm{W}\bm{\Sigma}^{-1}\bm{W}^{T}\). Let \(\mathcal{Z}\) denote the subset of \(\mathbb{R}^n\) where \(\bm{v}\) lives, that is, \(\mathcal{Z}\) is the Cartesian product of \(n\) half (positive or nonpositive) lines, where the \(i\)th component is \((0, \infty)\) (if \(y_i =1\)) or \((-\infty, 0]\) (if \(y_i =0\)). Let \(G\) be the unimodular multiplicative group on \(\mathbb{R}_+\) with Haar measure \(\nu(dg) = dg/g\), where \(dg\) is the Lebesgue measure on \(\mathbb{R}_+\). For constructing an efficient extra step, as in, we let the group \(G\) act on \(\mathcal{Z} \times \mathbb{R}_+^r\) through a group action \(T(\bm{v}, \bm{\tau}) = (g\bm{v},\bm{\tau}) = (gv_1, gv_2, \dots, gv_n, \bm{\tau})\). With the group action defined this way, it can be shown that the Lebesgue measure on \(\mathcal{Z} \times \mathbb{R}_+^r\) is relatively left invariant with multiplier \(\chi (g) = g^n\). Following, consider a probability density function \(\vartheta(g)\) on \(G\) where \[\vartheta\left(g\right) dg \propto \pi\left(g\bm{v}, \bm{\tau}|\bm{y}\right) \chi \left( g \right) \nu(dg) \propto g^{n-1}\exp\left\{-\frac{1}{2}g^{2}\bm{v}^{T} M_1 \bm{v} \right\} dg. \label{eq_g1}\] Since propriety of the posterior density [\[eq_joint_improper\]](#eq_joint_improper){reference-type="eqref" reference="eq_joint_improper"} implies that \(\pi(\bm{v}, \bm{\tau}|\bm{y})\) is a valid density, \(\bm{v}^TM_1\bm{v}\) can be zero only on a set of measure 0 (in \(\bm{v}\)). Thus given \((\bm{v},\bm{\tau})\), \(\vartheta\left(g\right)\) is a valid density. From, it follows that the transition \((\bm{v}, \bm{\tau}) \rightarrow (\bm{v}', \bm{\tau}) \equiv T(\bm{v}, \bm{\tau}) = (g\bm{v},\bm{\tau})\) where \(g \sim \vartheta(g)\), is reversible with respect to \(\pi(\bm{v},\bm{\tau}|\bm{y})\) defined in [\[eq:margvtau1\]](#eq:margvtau1){reference-type="eqref" reference="eq:margvtau1"}. Given \(\bm{\eta}^{(m)}\), below are the three steps involved in the \((m+1)\)st iteration of the Haar PX-DA algorithm to move to the new state \(\bm{\eta}^{(m+1)}\).
The extra step of Algorithm [\[algorithm_g1\]](#algorithm_g1){reference-type="ref" reference="algorithm_g1"} is a single draw from the univariate density \(\vartheta(g)\), which is easy to sample from. Thus, the computational burden, per iteration, for the Haar PX-DA algorithm is similar to that of the block Gibbs sampler described in section [3](#sec_twoblock){reference-type="ref" reference="sec_twoblock"}. Two other Haar PX-DA algorithms can be constructed by using group actions \(T_1(\bm{v}, \bm{\tau}) = (\bm{v}, g\bm{\tau})\) and \(T_2(\bm{v}, \bm{\tau}) = (g\bm{v}, g\bm{\tau})\). However, the corresponding \(\vartheta(g)\)'s are not easy to sample from, thus we do not consider them here.
# Discussion {#sec_conclusion}
We develop a two-block Gibbs sampler for the Bayesian probit linear mixed models under improper priors. The block Gibbs algorithm samples the fixed effects and the random effects jointly. We prove the geometric ergodicity of the two-block Gibbs sampler, which guarantees the existence of central limit theorems for MCMC estimators under a finite second moment condition. We also propose the corresponding Haar PX-DA algorithm. The Haar PX-DA algorithm not only improve the efficiency of the Gibbs sampler, but also inherit their geometric convergence properties. Another popular link function is the logit link function. proposed a DA algorithm for the logistic regression model. proved the uniform ergodicity of this DA algorithm. As mentioned in, their DA algorithm can be extended to the logistic linear mixed model. However, the convergence properties of the corresponding Markov chain have not been studied, and can be a topic for future research. Another future project can be deriving similar extensions of the results in for proving geometric convergence of Gibbs samplers for robit linear mixed models. | {'timestamp': '2018-11-26T02:13:32', 'yymm': '1706', 'arxiv_id': '1706.01846', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01846'} |
null | null |
null | null |
# Introduction {#sec:intro}
Crystals exposing a face bearing a net charge are intrinsically unstable if, in addition, the unit cell also has a net dipole moment perpendicular to the surface. Such a termination is referred to as a type III polar surface in the classification by Tasker. Tasker explained the instability of type III surfaces by showing that the energy diverges with increasing thickness of the crystal (polar catastrophe). Yet, surfaces with type III orientations do occur in nature. The electrostatic instability is avoided by the accumulation of compensating charge which cancels the dipole moment. Various compensating mechanisms have been observed or suggested. The dominant mechanisms, as reviewed by Noguera, are a change of surface composition (non-stoichiometric reconstruction), adsorption of charged species, and electronic reconstruction (charging of gap or defect states). The review of Ref. is restricted to oxide materials. It was updated in 2008 in collaboration with Goniakowski and again in 2013, now also including nano objects, such as thin films. Thin films are of special interest because some structures can exhibit an unreconstructed polar surface. Because of their small size they can sustain the polarization field driving the instability in larger systems. Polar surfaces are clearly a challenge for computational methods. The problem of how to calculate the total energy of model systems was first addressed as a theoretical exercise in the summation of electrostatic interactions arising from an array of point charges which was followed later by studies involving electronic structure calculations. The favourite model systems are: MgO (111)(rocksalt), free standing or deposited on a metal substrate;, NiO(111)(rocksalt), and ZnO, either the (111) zinc blende or (0001) (wurtzite) surface. More complex surfaces that have been modelled include tetragonal ZrO\(_2\)(110) and \(\alpha\)-\(\mathrm{Al}_2\mathrm{O}_3 (0001)\). More recently ternary oxides have attracted attention, in particular polar terminations of LaTiO\(_3\) and interfaces with SrTiO\(_3\). Atomistic model systems necessarily have a slab geometry of limited width. Electric fields in the solid are permitted. Similar to the thin films of experiment, the finite slab width can therefore mask a polar catastrophe. In this contribution we will outline a finite field approach adjusting the electrostatics. The method is inspired by our work on the atomistic modelling of an electric double layer (EDL) formed by a solid in contact with an aqueous solution (electrolyte). Finite electric fields penetrating the solid are again a serious concern if the solid is an insulating mineral. The surface charge is of chemical origin: the result of exchange (adsorption or desorption) with an ionic solution (electrolyte). The surface charge is compensated by a zone of excess ionic charge on the electrolyte side of the interface. In a macroscopic, semi-infinite solid the net charge in the EDL is zero. However this basic rule of EDL theory can be violated if the system is represented by a slab with surfaces of opposite charge by creating an internal electric field. The residual electric field manifests as a finite, net charge in the EDL. The surface charge is not properly screened by the mobile ions of the solution, instead the slab acts as a nano-capacitor. The nanocapacitor effect, while of interest by itself, must be regarded as a finite system size error if the aim is to simulate an interface between a semi-infinite insulator and an electrolyte. The magnitude of the error was investigated in Ref. for the familiar simple point charge (SPC) model of an aqueous NaCl solution confined between two oppositely charged walls. The walls separating the solution from vacuum were only an atomic diameter thick. Since full 3D periodic boundary conditions (PBC) were applied, moving the supercell over half its length generates a different perspective of the same system. Now the vacuum is in the middle and the system can be viewed as parallel plate capacitor. Even for gaps as large as 20 to 100 Å, the deficit charge in the EDL turns out to be a significant fraction of the surface charge (up to 20% at 20 Å). The results of Ref. were interpreted in terms of a simple continuum (Stern) model of the two EDL's. Within this model the missing EDL charge was found to be linearly correlated with the electric field in the insulator (the vacuum space). This observation suggested that charge neutrality of the EDL can be restored by applying an external bias field of the correct magnitude to cancel the field in the insulator. We verified that this is indeed the case, applying a classical molecular dynamics (MD) version of the finite field methods developed by Stengel et al. A further prediction of the continuum model was that the compensating field, termed the field of zero net charge (ZNC), is proportional to the inverse capacitance of the EDL. This relation was also confirmed by the atomistic simulation and used to compute the capacitance. In the present contribution the finite field methodology of Ref. is extended to polar interfaces. The idea is again that application of an appropriate external electric field should cancel the internal field associated with the polarization. In this first application the configuration of the ions in the solid is strictly fixed. Only the ions and water molecules in the electrolyte move and should therefore screen the polarization. The question is thus whether the excess charge supplied by the electrolyte can play the role of the compensating charge density in the theory of polar surfaces. The Tasker rule makes a very precise statement about the compensating charge which should be relatively easy to verify in a simple classical point charge model as will be used here.
# Model and method {#method}
## Point charge model for the NaCl-electrolyte interface {#sec:spc}
The main objective of this study is validation of the finite field method for electrochemical interfaces of polar surfaces. We opted therefore for the simplest of model systems, a slice of NaCl crystal of (111) orientation. The slab is terminated on one side by a Na\(^+\) plane and the other side by a Cl\(^-\) plane and is kept rigidly in the bulk rock salt geometry. The solid is in contact with a high concentration aqueous NaCl solution. Fig. [\[fig:renders\]](#fig:renders){reference-type="ref" reference="fig:renders"}a shows an instantaneous configuration sampled from the MD simulation. The classical force model is the same as used in Ref. (for further detail see section [2.3](#sec:md){reference-type="ref" reference="sec:md"}). The theory (Tasker rule) says that the surface charge density required to cancel the dipole generated by cleaving a rocksalt structure along a (111) plane is \(-\sigma_0/2\), where \(\sigma_0\) is the surface charge density of the terminating plane. A crystal with this solid/electrolyte interface should be stable and have no net internal electric field. The procedure is therefore the same as in Ref. . The applied bias field is varied until the internal Maxwell field in the slab is on average zero. We computed the charge imbalance (space charge) in the electrolyte in contact with the crystal face and checked whether it has the theoretical value of \(-\sigma_0/2\).
## Stern model for the polar surface-electrolyte interface {#sec:stern}
Anticipating our results we found that the Tasker half surface charge density rule for the unreconstructed rocksalt (111) termination is indeed satisfied for our model. This raises questions about the electrostatics of the polar surface/electrolyte interface. What are the differences compared to the regular charge neutral EDL induced by the surface charge of a non-polar dielectric solid? What is the capacitance or how can we even define a capacitance? We will address this problem by generalizing the continuum Stern model of Ref. . The model is pictured in Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}. As before the electrolyte is partitioned in a proper ionic conductor (\(\epsilon=\infty\)) and two boundary layers of each of thickness \(l_{\mathrm{H}}\). The dielectric constant in a boundary layer, to which we will continue to refer as a "Helmholtz layer", is \(\epsilon_{\mathrm{H}}\). The Maxwell field in the conducting region is strictly zero while it can be finite in a Helmholtz layer. This field is denoted by \(E_{\mathrm{H}}\). The polarization of the electrolyte is represented by the surface charge density \(\pm \sigma\) of the planes separating the conductor and the Helmholtz layer. With the surface charge of the solid fixed, the electrolyte surface charge \(\sigma\) is the central variable in the model. To model the polar slab we follow Noguera, representing the solid by a succession of planes with alternating surface charge density \(\sigma_0\). The planes are a distance \(R\) apart. There are \(n+1\) of these planes dividing the solid up into \(n\) slices, where \(n\) is an odd number. The dielectric constant is homogeneous throughout the solid and will again be indicated by \(\epsilon_\text{d}\). The electric field is however not the same everywhere. Counting from the left in Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"} the field \(E_1\) in the first layer (\(i=1\)) is different from the field \(E_2\) in the next layer (\(i=2\)). Because of the strict alteration of the surface charges on the planes the field in all odd numbered regions is \(E_1\) and in the even numbered layers \(E_2\) (note the convention of the field directions in Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}). To represent a MD supercell the Stern model is periodically repeated in the normal direction. The length of the supercell is \(L\). In the finite field method of Ref. the periodic MD system is subject to an electric field \(\bar{E}\) using an extended Hamiltonian (see section [2.3](#sec:md){reference-type="ref" reference="sec:md"}). Note that \(\bar{E}\) is not an external field \(E_0\) but the average of the Maxwell field. The product \(V =-\bar{E}L\) can therefore be directly interpreted as the potential difference across the MD cell. The corresponding cell potential \(V\) is the sum of all potential differences over the uniform layers making up the system and we can write \[\label{eq:barE} \bar{E}L =-2 E_{\mathrm{H}}l_{\mathrm{H}} + n\bar{E}_{d}R,\] with the field in the electrolyte set to zero (see Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}). \(n \bar{E}_d R\) is (minus) the potential across the polar solid with average field \(\bar{E}_d\) given by \[\label{eq:Ed} \bar{E}_\text{d} =-\frac{n-1}{2n}E_{2}+\frac{n+1}{2n}E_{1}.\] The Maxwell equations for boundaries give the following dependencies: \[\begin{aligned}
\epsilon_{\mathrm{H}}E_{\mathrm{H}} + \epsilon_\text{d}E_{1} & = & 4\pi\sigma_{0} \label{eq:maxwell1} \\[4pt] \epsilon_{\mathrm{H}}E_{\mathrm{H}} & = & 4\pi\sigma \label{eq:maxwell2} \\[4pt] \epsilon_\text{d}E_{2} + \epsilon_\text{d}E_{1} & = & 4\pi\sigma_{0}. \label{eq:maxwell3} \end{aligned}\] Exchanging the fields in the right-hand side of Eq. [\[eq:barE\]](#eq:barE){reference-type="ref" reference="eq:barE"} for charge densities using Eqs. [\[eq:maxwell1\]](#eq:maxwell1){reference-type="ref" reference="eq:maxwell1"}, [\[eq:maxwell2\]](#eq:maxwell2){reference-type="ref" reference="eq:maxwell2"} and [\[eq:maxwell3\]](#eq:maxwell3){reference-type="ref" reference="eq:maxwell3"} we obtain an expression for \(\sigma\) in terms of \(\sigma_0\) and \(\bar{E}\), \[\label{eq:sigma_flat} \sigma = \left( \frac{n+1}{2} \frac{\sigma_{0}}{C_\text{d}}-\bar{E}L \right) \left( \frac{2}{C_{\mathrm{H}}}+\frac{n}{C_\text{d}} \right)^{-1},\] where \(C_\text{d}= \epsilon_\text{d}/(4\pi R)\) and \(C_{\mathrm{H}}= \epsilon_{\mathrm{H}}/(4\pi l_{\mathrm{H}})\). The parameters \(C_\text{d}\) and \(C_{\mathrm{H}}\) have the familiar form of the capacitance of a parallel plate capacitor and will be interpreted as such. The surface charge \(\sigma_0\) is fixed. The Maxwell field \(\bar{E}\) is the only control parameter in the model. \(\bar{E}\) can be varied until the average field \(\bar{E}_d\) within the crystal is zero. In this state, referred to as the point of "Compensating Net Charge" (CNC), the two interfaces of the slab are decoupled, and so the infamous finite size error is removed. Setting \(\bar{E}_d = 0\) in Eq. [\[eq:barE\]](#eq:barE){reference-type="ref" reference="eq:barE"} and inserting into Eq. [\[eq:sigma_flat\]](#eq:sigma_flat){reference-type="ref" reference="eq:sigma_flat"} using Eq. [\[eq:maxwell2\]](#eq:maxwell2){reference-type="ref" reference="eq:maxwell2"} gives the compensating charge provided by the electrolyte: \[\label{eq:sigma_n} \sigma = \frac{n+1}{2n}\sigma_{0}.\] Indeed in the limit \(n\rightarrow\infty\), the prefactor of equation [\[eq:sigma_n\]](#eq:sigma_n){reference-type="ref" reference="eq:sigma_n"} tends towards \(\frac{1}{2}\) in agreement with the Tasker rule for a rocksalt(111) polar surface. The compensating charge is determined by the surface charge only and is independent of all other structural parameters. The \(\bar{E}=\bar{E}_{\mathrm{CNC}}\) field plays the same role as the field of zero net charge \(\bar{E}_{\mathrm{ZNC}}\) of Ref. . At \(\bar{E} = \bar{E}_{\mathrm{ZNC}}\) the field in the dielectric slab is zero restoring net charge neutrality to the EDL. Moreover the Stern model led to a simple relation between the capacitance of the EDL and \(\bar{E}_{\mathrm{ZNC}}\) (Eq. 37 of Ref. ) \[\label{eq:cHedl} C_{\mathrm{H}}=\frac{2\sigma_{0}}{\bar{E}_{\mathrm{ZNC}}L}\] Pursuing the parallel with the "regular" EDL of Ref. further, we can similarly express the capacitance of the Helmholtz layer in terms of \(\bar{E}_{\mathrm{CNC}}\). Combining Eq. [\[eq:sigma_flat\]](#eq:sigma_flat){reference-type="ref" reference="eq:sigma_flat"} and [\[eq:sigma_n\]](#eq:sigma_n){reference-type="ref" reference="eq:sigma_n"}, equivalent to imposing a potential \(-\bar{E}_{\mathrm{CNC}}L\), we find \[\label{eq:cH} C_{\mathrm{H}}=\frac{2\sigma_{0}}{\bar{E}_{\mathrm{CNC}}L}\frac{n+1}{2n}.\] Recall that the factor 2 multiplying \(\sigma_0\) in Eqs. [\[eq:cH\]](#eq:cH){reference-type="ref" reference="eq:cH"} and [\[eq:cHedl\]](#eq:cHedl){reference-type="ref" reference="eq:cHedl"} is there because \(-L\bar{E}\) is the potential over a pair of EDL's in series each with a capacitance \(C_{\mathrm{H}}\). The status of \(C_{\mathrm{H}}\) as the Helmholtz capacitance of the polar surface/electrolyte interface may look ambiguous because the response charge of the electrolyte is only half the surface charge. It is therefore of interest to compare the capacitance for the polar NaCl \((111)\) surface as defined by Eq. [\[eq:cH\]](#eq:cH){reference-type="ref" reference="eq:cH"} to the capacitance obtained by charging a non-polar surface. The obvious candidate is the \((100)\) surface. The surface was charged by slightly enhancing the positively (negatively) charged ions compared to the counter charge. The expression for \(C_{\mathrm{H}}\) is given by Eq. [\[eq:cH\]](#eq:cH){reference-type="ref" reference="eq:cH"} leaving out the factor \((n+1)/(2n)\) consistent with Eq. [\[eq:cHedl\]](#eq:cHedl){reference-type="ref" reference="eq:cHedl"}.
## Hamiltonian and finite electric field molecular dynamics {#sec:md}
All simulations were performed under ambient conditions using a modified version of the GROMACS package. The water model is SPC/Extended. The SPC model of aqueous Na\(^+\) and Cl\(^-\) was taken from Ref. and has been validated for high salt concentrations. (see also the review by Nezbeda et al. ). Identical force field parameters were used for the interactions with the ions in the (rigid) NaCl crystal. The supercell cross-sectional area were 2.200 and 2.545 nm\(^2\) for the (111) and (100) orientation respectively. The corresponding distance between the charged planes of the (111) slab, referred to as \(R\) in Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}, is \(R=1.63\) Å. The MD cell lengths were adjusted to keep variations in solvent properties to a minimum. All supercells contained 20 aqueous NaCl ion pairs in \(\sim\) 600 waters, evenly dispersed in the initial configuration. This particular concentration was chosen such that the electrolyte remains a good ionic conductor after the formation of EDLs has reduced the concentration (the total number of ions is fixed). The MD parameters were as follows: the NVT ensemble was employed, and a temperature of 298 K mainted by Nosé-Hoover thermostat with a coupling constant of 0.4 ps; the timestep was 2 fs and the simulations were run for a total of 1 ns; the electrostatics were computed using 4\(^{th}\) order Particle Mesh Ewald (PME) summation with a Fourier spacing of 0.6 Å and a real-space cutoff of 6.5 Å. The first 200 ps were discarded as equilibration time. The system was considered equilibrated if the electrostatic potential in the bulk electrolyte was flat. Ref investigated the equilibration time for \(C_{\mathrm{H}}\) and the choice of 200 ps achieves this level of convergence in the present work. The modification of the GROMACS package concerns the implementation of the constant field Hamiltonian method according Vanderbilt and co-workers. Forces and energies are derived from the extended Hamiltonian \[\label{fhmd} H_E\left( v, \bar{E} \right) = H_{\mathrm{PBC}}(v)-\mathit{\Omega} \bar{E} P(v),\] where \(H_{\textrm{PBC}}(v)\) is the Hamiltonian as defined by the SPC model. \(v= (\mathbf{r}^N ,\mathbf{p}^N)\) stands for the collective momenta and position coordinates of the \(N\) particles in the system. The subscript PBC indicates that the electrostatic energies and forces are computed using standard Ewald summation ("tin foil" boundary conditions). \(\mathit{\Omega}\) is the volume of the MD cell. \(P\) is the polarization perpendicular to the crystal slab with \(\bar{E}\) the magnitude of the electric field. \(P\) is computed from the total dipole moment of the supercell including the contribution from the ions in the solution and the solid. Typical values of \(\bar{E}\) applied to our model system are in the order of 1 V/nm. Finite electric field Hamiltonians of the form of Eq. [\[fhmd\]](#fhmd){reference-type="ref" reference="fhmd"} are familiar in classical MD. As already pointed out by Yeh and Berkowitz, the electric field in the dipole coupling term \(-\mathit{\Omega} \bar{E} P\), when combined with Ewald summation, must be interpreted as the average of the Maxwell field rather than the applied external field. In Ref. this feature, specific to standard Ewald summation, is given a firm thermodynamic foundation (see also Ref. ). A further important point is that including ions in the polarization makes \(P\) a multi-valued function depending on the supercell boundary. This is a central theme in the modern theory of polarization which applies to electronic polarization as well as classical point charge systems. This issue becomes critical for a supercell of the geometry of Fig. [\[fig:renders\]](#fig:renders){reference-type="ref" reference="fig:renders"}. The mobile ions can cross the cell boundaries. When this happens the ions must be followed out of the cell in the calculation of the polarization. In classical simulation, this definition of polarization was introduced in the context of the MD study of electrolytic solutions and is generally referred to as itinerant polarization. Note that the forces derived from the Hamiltonian Eq. [\[fhmd\]](#fhmd){reference-type="ref" reference="fhmd"} are not affected by ambiguities in the definition of polarization (see Ref. for a more detailed discussion).
# Results {#sec:results}
## Finding the field of compensating net charge (CNC) {#sec:cnc}
The first step is locating the field of compensating net charge. In the state of CNC, \(\bar{E}_\text{d}=0\) by definition. The change \(\Delta \phi_\text{d}\) in the electrostatic potential over the length of the crystal must therefore be zero. \(\Delta \phi_\text{d}\) was computed from the electrostatic potential profile \(\phi(z)\). As is evident from Fig. [\[fig:graphspot\]](#fig:graphspot){reference-type="ref" reference="fig:graphspot"}(a) the potential profile, while flat in the electrolyte, shows a saw tooth pattern in the crystal reflecting the alternating charge of the (111) crystal planes. In contrast, the potential profile in the non-polar slab (Fig. [\[fig:graphspot\]](#fig:graphspot){reference-type="ref" reference="fig:graphspot"}(b)) is smooth because (100) planes are net neutral. In a CNC state (\(\bar{E}_\text{d}=0\)), from which the configuration of Fig. [\[fig:graphspot\]](#fig:graphspot){reference-type="ref" reference="fig:graphspot"}(b) was sampled, the potential is therefore constant (\(\Delta \phi_\text{d} = 0\)). It is perhaps instructive to analyze the polar profile of Fig. [\[fig:graphspot\]](#fig:graphspot){reference-type="ref" reference="fig:graphspot"}(a) in somewhat more detail because its appearance may at first seem inconsistent with the condition of vanishing average internal electric field required at CNC. Referring back to Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"} we note that the period of the modulation of the potential in Fig. [\[fig:graphspot\]](#fig:graphspot){reference-type="ref" reference="fig:graphspot"}(a) is \(2R\). However, the width of the slab is \(nR\) where \(n\) is an odd integer. The CNC field therefore aligns a maximum of the saw tooth at one face with a minimum at the opposite face as indicated by the red dashed line in the figure. This can only be achieved by canting the sequence of maxima (minima) which is the feature that stands out in Fig. [\[fig:graphspot\]](#fig:graphspot){reference-type="ref" reference="fig:graphspot"}(a). \(\Delta \phi_\text{d}\) was tracked against the applied field \(\bar{E}\) to locate the point of CNC for increasing width of the crystal. The result is shown in Fig. [\[fig:montage\]](#fig:montage){reference-type="ref" reference="fig:montage"}(c). The width of the slab is represented by the number \(n\) of capacitors in series (see Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}). The figure suggests that a minimum of \(n=15\) layers (16 planes) is needed to reach convergence. This corresponds to a slab width of \(25\) Å. In the search for the CNC it was observed that the response of charge and potential to changes in the field was linear. The exception is the terminal n=1 system. This is likely due to dielectric saturation at the high CNC field for n=1 as suggested by Fig. [\[fig:montage\]](#fig:montage){reference-type="ref" reference="fig:montage"}(c). Therefore, the values for n=1 are obtained by extrapolation from the low-field, linear regime. Slab width is of course always a critical parameter in periodic models of interfaces. Solid NaCl/aqueous NaCl interfaces are popular model systems and the question of size dependence has been investigated in great detail in calculations of solubility from the direct equilibrium between solid and solution. Using an interaction model identical to the one used here, Espinoza and coworkers found in a careful study that the solubility is essentially converged for slabs of a width larger than 40 Å. This number exceeds but is still comparable to the \(25\) Å inferred from Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}. Considering however, that the orientation of the slab used in Ref. is the nonpolar \((100)\) surface, the similarity between the finite size effects is perhaps somewhat surprising. In this context we should also point out that the size effect for our polar interface seems to be captured to great accuracy by an analytical expression (Eq. [\[eq:sigma_n\]](#eq:sigma_n){reference-type="ref" reference="eq:sigma_n"}) specified by only a single parameter, the surface charge density \(\sigma_0\) of the polar \((111)\) surface for which there is no direct counter part in the dissolution of a \((100)\) interface. On the other hand the consistency in finite size effects can also be interpreted as an indication that surface charge and polarity play a role in the dissolution of non-polar interfaces.
## Compensating electrolyte charge and capacitance {#sec:edlcap}
In the simple Stern model of Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}, the charge induced in the electrolyte is represented as a surface charge density \(\sigma\) at the sharp interface between the electrolyte and the Helmholtz layer. In the atomistic model system of Fig. [\[fig:renders\]](#fig:renders){reference-type="ref" reference="fig:renders"}, the interface is more diffuse. We have estimated the compensating charge as the integral of the excess charge density ("space charge") of the electrolyte in contact with the surface. This procedure is outlined in more detail in Ref. . The excess charge per unit area is identified with the \(\sigma\) of the Stern model. The results are plotted in Fig. [\[fig:montage\]](#fig:montage){reference-type="ref" reference="fig:montage"}(a) where we have represented the estimates of \(\sigma\) as total surface charges \(A \sigma\) with \(A\) the MD cell cross section. The surface charge of a NaCl (111) plane in our model system is \(A \sigma_0 = 16e\). The theoretical compensating charge (Tasker rule) is therefore \(8e\). Fig. [\[fig:montage\]](#fig:montage){reference-type="ref" reference="fig:montage"}(a) shows that the compensating charge \(A \sigma\) approaches \(8e\) when the number of layers of polar solid gets large enough. Therefore, our simulation confirms that the charge imbalance is in accord with the theoretical value. Indeed, figure [\[fig:montage\]](#fig:montage){reference-type="ref" reference="fig:montage"}(b) shows excellent agreement with Eq. [\[eq:sigma_n\]](#eq:sigma_n){reference-type="ref" reference="eq:sigma_n"} derived from the continuum model Fig. [\[fig:stern\]](#fig:stern){reference-type="ref" reference="fig:stern"}. This also gives us confidence that Eq. [\[eq:cH\]](#eq:cH){reference-type="ref" reference="eq:cH"} can be used to extract a value for the Helmholtz capacitance from simulations. Indeed, Fig. [\[fig:montage\]](#fig:montage){reference-type="ref" reference="fig:montage"}(d) shows the desired linear (n+1)/2n dependence of Eq. [\[eq:cH\]](#eq:cH){reference-type="ref" reference="eq:cH"}. The slopes gives an estimation of \(C_\text{H}\) of 8.23 \(\mu\)Fcm\(^{-2}\) for the polar (111) surface. 8.2 \(\mu\)Fcm\(^{-2}\) is a relatively modest capacitance, not much larger than the 4.4 \(\mu\)Fcm\(^{-2}\) we found in Ref. for the EDL formed by a 1.4 M aqueous NaCl electrolyte confined between two walls of uniform opposite surface charge densities. While our procedure of a fractional increase of the charge of all ions of one species (Na\(^+\) or Cl\(^-\)) is nonphysical, we argue that the electrostatics of an EDL formed by this system is similar to that of the "regular" EDL of Ref. . Accordingly, the Helmholtz capacitance can be obtained from Eq. [\[eq:cHedl\]](#eq:cHedl){reference-type="ref" reference="eq:cHedl"} or equivalently from Eq. [\[eq:cH\]](#eq:cH){reference-type="ref" reference="eq:cH"} by leaving out the \((n+1)/2n\) factor. The EDL of the non-polar surface is charge balanced for vanishing internal field. The CNC and ZNC are interchangeable in this case. The result of the applied voltage at CNC as a function of the surface charge for the non-polar (100) surface is shown in Fig. [\[fig:chargevznc\]](#fig:chargevznc){reference-type="ref" reference="fig:chargevznc"}. The slope yields the Helmholtz capacitance \(C_\text{H}\) with a value 4.23 \(\mu\)Fcm\(^{-2}\), effectively identical to the capacitance obtained in Ref. .
## Electric double layer structure {#sec:edlstruc}
The near 2 to 1 ratio for the capacitance of the polar (111) and non-polar (100) surfaces may not be a coincidence. A possible explanation for this behaviour can be seen in the double layer structure of Fig. [\[fig:densityplot\]](#fig:densityplot){reference-type="ref" reference="fig:densityplot"}. For the (111) case, the ions of the double layer are adsorbed to the surface at a distance of \(\sim\)`<!-- -->`{=html}1.5 Å, shedding their inner solvation shell. The formation of contact ion pairs is particularly noteworthy for the Na\(^{+}\) ions, which are traditionally thought to have an almost unbreakable solvation shell. In the (100) case, at the lowest charge of 2e, there is no driving force for such a dehydration, and the first peak in the density occurs at the larger distance of 2.8 Å, followed by a small secondary peak further out. As the surface charge is increased, the profile shifts towards the surface. Even for a (100) surface charge as high as the compensating charge for the (111) polar surface (\(8e\)), the structural difference between EDLs is apparent. The capacitance of a compact (Helmholtz) EDL is inversely proportional to its width, therefore, the 2 to 1 ratio in \(C_{\mathrm{H}}\) between the surfaces is roughly consistent with the relative positions of the counter ions in Fig. [\[fig:densityplot\]](#fig:densityplot){reference-type="ref" reference="fig:densityplot"}.
# Conclusion and outlook
Our calculations show that the excess charge in a high concentration NaCl aqueous solution adjacent to a rigid NaCl (111) surface complies to the Tasker rule for a polar surface of this geometry (half the surface charge). This was to be expected. The Tasker rule is based on general considerations involving geometry and electrostatics. It would have been a surprise if this rule would not hold for polar surface electrolyte interfaces. The results of the present study are therefore intended as further validation of our finite field method for the simulation of electric double layers under full periodic boundary conditions. This method was introduced in Ref. and applied to a conventional electric double layer for which the electrolyte counter charge equals the opposite of the surface charge (zero net charge). Reproducing the half charge predicted by the Tasker rule for a model polar surface seemed to us a separate challenge which was taken up in the calculation reported here. The key function of the applied external field was to compensate for the internal electric field in polar model slabs. Slabs of a width accessible to molecular simulation can sustain an internal field leading to violation of the Tasker rule for polar surfaces of semi-infinite crystals. The fact that the electrolyte is an ionic conductor is essential. This enforces a zero macroscopic field in the bulk region of the electrolyte whatever the magnitude of the polarization of the slab or the applied field. The external field can therefore be adjusted to cancel the internal field in the solid without inducing a field in the electrolyte. The model system in this feasibility study was deliberately kept as simple as possible. In particular the structure of the solid slab was constrained to be rigid. This meant that a number of interesting issues could not be addressed. One most important question is the competition with non-stoichiometric reconstructions observed for vacuum surfaces. One of the candidate structures for the rocksalt (111) surface is the so-called octopolar reconstruction suggested by Wolf (see also Noguera). Dissolution of NaCl is a facile process interfering with reconstruction and the stability of polar surfaces is usually studied for more robust rocksalt crystals such as MgO and NiO. However, under saturation the NaCl surface is in equilibrium with its aqueous solution and it has been proven to be feasible to stabilize a NaCl(111)/NaCl(aq) interface under appropriate thermodynamic conditions. The force field used in the present study (the Joung-Cheetham model of Ref. ) is also suitable to carry out a simulation of a solid/solution equilibrium. After years of hard work there seems to be now a consensus in the computational literature regarding the solubility of this model (see the review of Ref. ). This is below the experimental value, but above the 2 M concentration used in the present study for the solution phase. This ensures that there is no net tendency of the crystal to grow. These observations apply to non-polar surfaces (see however the results of Ref. for a spherical piece of crystal). It would therefore be of interest to repeat these calculations for a polar interface. Finally we return to the capacitance we computed in section [3.2](#sec:edlcap){reference-type="ref" reference="sec:edlcap"} for the polar surface electrolyte interface. The possibility of experimental realization of an unreconstructed polar surface electrolyte interface raises the question of the interpretation of this capacitance and how it might be observed. While the structure of the "double layer" should be a least in principle verifiable by experiment, the status of the corresponding capacitance is less clear. Capacitance is well defined for a nanoslab and should also be accessible to experiment. It would be the capacitance of the nanocapacitor formed by the solid slab with the electrolyte on either side acting as the two electrodes. But how to determine the capacitance of the interface of a non-conducting (insulator) semi-infinite crystal with a polar termination? The capacitance probably enters experimentally-measurable quantities only indirectly, such as in adsorption (complexation) energies of charged species. Maybe the double layer will also show electrokinetic signatures, such as a finite zeta-potential. This rather technical report is not the place to address these questions. This will have to be resolved in future investigations. In conclusion we reiterate the observation by Noguera that the value of the compensating charge is a result of long range electrostatics and should remain the same even if the electronic structure is taken into account. We therefore anticipate that this method can enable us to study the interaction of an electrolyte with more complex and realistic polar surfaces, possibly even treated by electronic structure calculation. | {'timestamp': '2018-09-05T02:28:23', 'yymm': '1706', 'arxiv_id': '1706.01947', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01947'} |
# Introduction {#sec:intro}
Credit card frauds, a concept included in the wider notion of financial frauds, is a topic attracting an increasing attention from the scientific community. This is due, on the one hand, to the raising costs that they generate for the system, reaching billions of dollars in yearly losses and a percentage loss of revenues equal to the \(1.4\%\) of online payments. On the other hand, credit card frauds have important social consequences and ramifications, as they support organised crime, terrorism funding and international narcotics trafficking-see for a complete review. Detecting unauthorised credit card transactions is an extremely complex problem, as features are seldom useful if taken individually. To illustrate, a large transaction is not *prima facie* suspicious, unless it is performed at usual times (*e.g.* at night) or in an unusual store (*i.e.* a store never visited before by the card owner, located in a different city, *etc.*). When different features have to be combined in non-trivial ways, the customary solution is to resort to data mining, a sub-field of computer science dealing with the automatic discovery of patterns in data sets. Most of the data mining models to detect credit card frauds are based on artificial neural networks (ANN), a model inspired on the structural aspects of biological neural networks, and in which a set of nodes process the input signal by interacting between them. This does not come as a surprise, as ANN are able to extract complex non-linear patterns from data, with almost no hypotheses on the underlying structure. ANN yielded good results in credit card classification tasks, as for instance in-see for reviews. While data mining algorithms are able to detect hidden patterns in data, they usually lack the capacity of synthesising metrics describing the global structure created by the interactions between the different features. In recent years, the use of complex networks theory has been proposed as a way of overcoming this limitation. Complex networks are a statistical-mechanics understanding of the classical graph theory, aimed at describing and characterising the structure of complex systems. The interaction between network theory and data mining is bidirectional: the former can be used to synthesise high-level features to be fed into a classification problem; while the latter can endow networks with an objective way of validating results-see for a complete review. More specifically, complex networks and data mining can be integrated as complementary tools in order to extract, synthesise and create new representations of a data source, with the aim of, for instance, discover new hidden patterns in a complex structure. The appropriate integration of complex network metrics can result in improved classification rates with respect to classical data mining algorithms and, reciprocally, there are many situations in which data mining can be used to solve important issues in complex network theory and applications. In this contribution we explore the possibility of using complex networks as a way of improving credit card fraud detection. Specifically, networks are used to synthesise complex features representing card transactions, relying on the recently proposed approach of *parenclitic networks* (Section [2](#sec:methods){reference-type="ref" reference="sec:methods"}). Afterwards, their relevance is evaluated by means of a large dataset of real transactions, by comparing the yielded increase in the classification score when compared to the use of a standard ANN algorithm (Section [3](#sec:results){reference-type="ref" reference="sec:results"}). We additionally show that the combined data mining / complex networks approach is able to outperform a commercial system in some specific situations.
# Methods {#sec:methods}
In this section we present the main tools that are going to be used for the classification of credit card transactions between licit and illicit. Given a credit card transaction \(t_i\) with features \(f_{i1},\cdots,f_{ik}\), the problem entails detecting if it is illicit or not from its features and the knowledge obtained from an historical training dataset-what is known as a supervised learning problem. From a mathematical point of view, we have to model a function \(H:\mathbb{R}^k\longrightarrow\mathbb{R}\) and find \(\delta>0\) such that if \(|H(f_{i1},\cdots,f_{ik})|\le \delta\), then \(t_i\) is not illicit. Note that, while there are multiple types of illicit patterns, such aspect is here not considered, in that any suspicious transaction is considered as a potential fraudulent one. We firstly introduce the concept of *parenclitic networks* in Section [2.1](#sec:parenclitic){reference-type="ref" reference="sec:parenclitic"}, a network reconstruction technique that allows highlighting the differences between one instance and a set of standard (*i.e.* baseline, or in this case licit) instances. We subsequently describe the real data set used for validation (Section [2.2](#sec:dataset){reference-type="ref" reference="sec:dataset"}), including the available raw features (Table [2](#tab:Features){reference-type="ref" reference="tab:Features"}); and the global classification model (Section [2.3](#sec:ClassModels){reference-type="ref" reference="sec:ClassModels"}).
## Parenclitic networks reconstruction {#sec:parenclitic}
As initially proposed in, one may hypothesise that the right classification of an observation does not only come from its features, but also from the structure of correlations between them. Following the mathematica formalism introduced before, if we consider the set \[L=\left\{(x_1,\cdots,x_k)\in\mathbb{R}^k;\enspace |H(x_1,\cdots,x_k)|\le \delta\right\}\subseteq\mathbb{R}^k,\] then \(L\) is a manifold in \(\mathbb{R}^k\) such that if we take a (new) transaction \(t\) with features \(t_1,\cdots,t_k\) such that \((t_1,\cdots,t_k)\notin L\), then \(t\) is considered as an illicit transaction. In general it is computationally impossible to obtain the set \(L\) directly from the training dataset, since it is a high dimensional problem. As an alternative, the parenclitic approach analyses the family of projections of \(L\) into 2-dimensional spaces corresponding to couples of features \((x_i,x_j)\) with \(1\le i\ne j\le k\). Hence, if we consider a training dataset with \(n\in\mathbb{N}\) transactions, each of them described by \(k\in\mathbb{N}\) (numeric) features, we can analyse up to \({k\choose 2}=k(k-1)/2\) two-dimensional projections of pairs of different features, each of them with up to \(n\) points in \(\mathbb{R}^2\). In order to quantify the correlation between pairs of features, the parenclitic approach proposes associating a network to each transaction with \(k\) nodes (as many as features considered) and the links measure the correlation between features. Hence the following pre-processing must be completed: for every two-dimensional projection of \(L\) given by a couple of features \((f_i,f_j)\) with \(1\le i\ne j\le k\), the correlation for the licit transactions in the training dataset is measured (by means of, for instance, a linear regression or other curve fitting techniques). For the shake of simplicity, we have here considered a linear regression, such that every pair of features \((f_i,f_j)\) with \(1\le i\ne j\le k\) yields a linear fitting between \(f_i\) and \(f_j\) for the licit transactions in the training dataset. Mathematically, this is represented by a linear equation of the form: \[r_{ij}:\,x_j=a_{ij}x_i+b_{ij}.\] Once these \({k\choose 2}\) linear regression lines are computed, a threshold \(\alpha>0\) is fixed. Given a new (*i.e.* not included in the training set) transaction \(t\) with features \(t_1,\cdots,t_k\), a network \(G=G(t)\) is associated to \(t\) as follows:
- \(G\) has \(k\) nodes \(1,\cdots,k\),
- For every pair of nodes \(1\le i\ne j\le k\) we compute \(w_ij\ge 0\) as the (euclidian) distance from \((t_i,t_j)\) to the line \(r_{ij}\) in \(\mathbb{R}^2\), i.e. \[w_{ij}=d((t_i,t_j),r_{ij}).\] As an alternative, the euclidian distance could be replaced by any pseudo-distance function in \(\mathbb{R}^2\). For the shake of simplicity, the euclidian distance will be used in this paper, but similar results can be obtained for other pseudo-distance functions.
- For every pair of nodes \(1\le i\ne j\le k\), the (undirected) link \((i,j)\) is in graph \(G\) if and only if \(w_{ij}\ge \alpha\).
Note that the parenclitic network \(G(t)\) summarises the couples of features whose correlation strongly differs from a typical licit transaction; the structure of this network thus contains valuable information about the (abnormal) correlation of features in the credit card transaction. Once this parenclitic network is computed, it is necessary to transform it in a set of features compatible with a data mining algorithm. Towards this end, several structural measures have been extracted, and will be considered as new features associated with the transaction (see next section for details). Among all possible structural measures that could be computed (see, for example, Ref. and references therein), those here selected are summarised in Table [1](#table:measures){reference-type="ref" reference="table:measures"}.
:::
## Classification models {#sec:ClassModels}
As previously introduced, in this contribution we are going to explore two different ways of detecting illicit credit card transactions: a classical data mining approach, and the introduction of features extracted from a network representation. In both cases, the process must follow some common steps: it is first necessary to extract the expected behavior, *i.e.* a set of features representing the typical legal and illegal transaction; for then building a model that learns from those features, and yields an expected classification for a new transaction not yet studied. Fig. [\[fig01\]](#fig01){reference-type="ref" reference="fig01"} depicts an overview of the whole process. It starts from the original data set, from which a set of raw features are extracted-as described in Section [2.2](#sec:dataset){reference-type="ref" reference="sec:dataset"} and listed in Tab. [2](#tab:Features){reference-type="ref" reference="tab:Features"}. The features corresponding to the licit transactions are then used to recover the normal relations, as described in Section [2.1](#sec:parenclitic){reference-type="ref" reference="sec:parenclitic"}, and to reconstruct the parenclitic networks of all transactions. These networks are then binarised, *i.e.* links with weight below a given threshold are deleted, and a set of topological metrics are extracted-see Table [\[tab:Topological\]](#tab:Topological){reference-type="ref" reference="tab:Topological"} for a complete list. Note that, at the end of this analysis, all transactions are described by \(15\) features: \(8\) coming from the raw data, and \(7\) from the network analysis. Artificial Neural Networks (ANNs), and specifically Muti-Layer Perceptrons (MLP) have been chosen as the final model for classifying new transactions. They are inspired by the structural aspects of biological neural networks, and are represented by a set of connected nodes in which each connection has a weight associated with it, and the network learns the classification function adjusting the node weights . The output of each artificial neuron \(j\) is defined by: \[f(W^T, x_j)=\sum^{n}_{i=1}{W_i x_i + b},\] \(W\) being the vector of weights, and \(f\) the sigmoid activation function: \[f(x)= \frac{1}{1 + \exp(-x)}.\] Following the standard configuration, neurons were organized in three layers: an input one, with a number of neurons equal to the input features; an intermediate, or hidden one, with ten neurons; and a final output layer comprising just one computational element. The training has been performed with the standard back-propagation algorithm . Finally, the reconstruction of the MLP models has been performed using the KNIME software. The evaluation of the classification efficiency has been performed using both sensitivity (also known as True Positive Rate-TPR) and Receiver Operating Characteristic (ROC) curves. These curves are created by plotting the True Positive Rate (TPR) against the False Positive Rate (FPR) at various threshold settings. ROC plots present the important advantage of showing the performance of the classification model for different sensitivity values. This is relevant for the problem at hand, as false positives are extremely expensive, *e.g.* in terms of the negative commercial image of the bank; conservative solutions are therefore usually preferred.
# Results {#sec:results}
As explained in Section [2.1](#sec:parenclitic){reference-type="ref" reference="sec:parenclitic"}, the parenclitic approach usually requires the definition of a threshold \(\alpha\), which is used to binarise the (initially weighted) networks. Instead of using an *a priori* approach, *i.e.* the definition of \(\alpha\) using expert judgement, we here tackle the problem indirectly, by following the procedure proposed in Ref.. Specifically, we optimise the network reconstruction by finding the link density (and hence the value of \(\alpha\)) that optimises the efficacy of the classification model. Fig. [\[fig02\]](#fig02){reference-type="ref" reference="fig02"} Left presents the evolution of the classification error (sensitivity or TPR) as a function of the considered link density, for three different scenarios: the use of only the raw features, as described in Tab. [2](#tab:Features){reference-type="ref" reference="tab:Features"} (solid black squares); the use of the features extracted from the parenclitic representation alone (hollow black circles); and the use of the combined sets of features (solid blue triangles). Note that, in the former case, the result is constant, as the original features are not affected by the binarisation process. In order to avoid overfitting, this classification has been performed on a balanced sub data set, composed of an equal number of legal and illegal transactions. Several conclusions can be drawn from Fig. [\[fig02\]](#fig02){reference-type="ref" reference="fig02"}. First of all, the features extracted from the parenclitic networks are not enough, alone, to reach a low classification error. This has to be expected: while important information can be codified in the interaction between raw features, some important clues may be hidden in the latter, *e.g.* abnormal transaction sizes or timings. At the same time, the addition of parenclitic features to the raw data set enhance the obtained results, with the error dropping from a \(19.2\%\) to a \(12.23\%\). This is further illustrated in Fig. [\[fig02\]](#fig02){reference-type="ref" reference="fig02"} Right, depicting the reduction in the classification error (in percentage) when considering only parenclitic features and the whole data set-note that, in the first case, the reduction is negative as the error increases. Finally, the best classification suggests that the optimal link density that should be considered is of a \(60\%\)-meaning that the \(40\%\) of links with less weight should be deleted. If Fig. [\[fig02\]](#fig02){reference-type="ref" reference="fig02"} is useful to detect the best link density for the analysis, it does not convey information about the real performance of the classification algorithm in an operational environment. For that, Fig. [\[fig03\]](#fig03){reference-type="ref" reference="fig03"} Left presents three ROC curves, corresponding to the use of raw (blue line), parenclitic (green line), and combined features (black line) as before. Note that results here presented correspond to the optimal link density of \(60\%\), as previously estimated. As previously discussed, the most interesting operational configuration is the one minimising the number of false positives, as this minimises the commercial costs of the organisation. The inset of Fig. [\[fig03\]](#fig03){reference-type="ref" reference="fig03"} thus shows the bottom left part of the curves. It can be appreciated that, after an initial part in which results are comparable, the addition of the parenclitic features slightly increases the number of true positives-note how the black line is above the blue one. Even though this may seem a negligible difference, it is worth noting that any improvement, however small, has a significant impact due to the large number of transactions managed by the system. Increasing the fraud detection rate by \(1\%\) would allow identify \(\approx 20.000\) new illicit transactions per year, or \(\approx 2\)M€ in saved costs. Fig. [\[fig03\]](#fig03){reference-type="ref" reference="fig03"} Right further presents four ROC curves calculated for different transaction sizes: all transactions (black line), and transactions above \(100\)€ (green line), \(400\)€ (red line) and \(1.600\)€ (blue line). Deleting small transactions results in an improvement of the detection efficiency-note how the green and red lines lay above the black one. Additionally, the proposed algorithm fails for large transactions; this does not come as a surprise, as the larger the size, the fewer the available instances, making training more challenging. If what previously presented illustrates that the use of a network representation can improve a fraud detection algorithm, it does not clarify how it ranks against a commercial system. As may be expected, the proposed algorithm is less efficient than the fraud score included in the original data set-see Fig. [\[fig04\]](#fig04){reference-type="ref" reference="fig04"} Left[^1]. Nevertheless, there are niches in which the opposite happens, the most important being the analysis of on-line transactions. Fig. [\[fig04\]](#fig04){reference-type="ref" reference="fig04"} Right depicts two ROC curves, respectively for the algorithm based on parenclitic networks (black line) and the commercial system (blue line), when only transactions realised through Internet are considered. While the commercial system clearly outperforms the proposed algorithm, with an Area Under the Curve (AUC) close to \(1.0\), the latter is slightly better for a low ratio of False Positive-as previously explained, the plane region most interesting for real operations.
# Conclusions
Complex networks and data mining models share more characteristics that what may *prima facie* appear, most notably having similar objectives: both aim at extracting information from (potentially complex) systems to ultimately generate new compact quantifiable representations. At the same time, they approach this common problem from two different approaches: the former by extracting and quantitatively evaluating the underlying structure, the latter by creating predictive models based on historical data. In this contribution we test the hypothesis that complex networks can be used as a way to improve data mining models, framed within the problem of detecting fraud instances in credit card transactions, providing a new example about how complex networks and data mining may be integrated as complementary tools in a synergistic manner in order to improve the classification rates obtained by classical data mining algorithms. Results confirm that features extracted from a network-based representation of data, leveraging on a recently proposed parenclitic approach, can play an important role: while not effective in themselves, such features can improve the score obtained by a standard ANN classification model. We further show how the resulting model is especially efficient in detecting frauds in some niches of operations, like medium-sized and on-line transactions. Finally, we illustrate as, in the latter case, the network-based model is able to yield better results than a commercial fraud detection system. All results have been obtained with a unique data set, comprising all transactions managed during two years by a major Spanish bank, and including more than \(180\) million operations.
[^1]: Due to confidentiality issues, the name and characteristics of the commercial fraud detection system cannot be included in this publication. | {'timestamp': '2017-06-08T02:01:53', 'yymm': '1706', 'arxiv_id': '1706.01953', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01953'} |
# Introduction
Innovative industries invest resources (e.g., money and time for research and development) to construct new systems and to improve the performance of the previously-deployed ones. To generate revenue and offset the cost of research, they ideally want to capitalize on their achievements. This is sometimes done by restricting the use of their ideas through patents or by hiding the features of their systems as trade secrets. When opting for trade secrets, reverse engineering techniques can be used by competitors to unravel their secrets. For instance, model identification tools can be utilized to identify a black-box system or to extract the parameters of a gray-box system. The gained information can be then used to reverse the financial gains. This motivates the use of methods that can render reverse-engineering techniques ineffective. Such methods, however, most often degrade the performance of the system. Therefore, a framework for balancing the need for preserving the trade secrets against maintaining the performance of the systems is required. In this paper, linear time-invariant discrete-time finite impulse response (FIR) system are considered. Specifically, the idea of adding noises to the input and output for increasing the error of model identification is explored. A bound on closed-loop performance degradation caused by the additive noise is enforced. An optimal filter for constructing the additive input and output noises that maximizes the identification error subject to maintaining the performance degradation below a threshold is constructed. This is done for both known and unknown input sequences. The former is useful to make the identification difficult for given inputs, such as the optimal experimental design in the model identification literature . The latter, which requires statistics of the input, can accommodate the belief of the designer on the reverse engineering techniques, e.g., a frequently used input for model identification purposes is a sequence of i.i.d.[^1] Gaussian noise . Finally, differential privacy framework is used for designing output additive noises that make the system identification difficult without any assumptions on the utilized inputs. In differential privacy literature, noises are added to the outcome of statistical queries from databases to preserve the privacy of individuals in the database . This framework was more recently used in dynamical systems . In differential privacy literature, most often, additive Laplace noises are used and the parameters of the noise are selected according to the sensitivity of the outcome to variations in the data (that should be kept private). However, weaker variants of differential privacy can be achieved by additive Gaussian noises. This is advantageous as adding Laplace noise can make the designer's task considerably more difficult (in terms of utilizing the outputs of the system), e.g., optimal state estimation when measurements are corrupted by Laplace noise results in non-linearities and memory issues . To the best of our knowledge, the differential privacy has not been explored in the context of preserving the privacy of dynamical systems with the aim of protecting the model as a trade secret. This has been explored thoroughly in one of the sections of the paper. In addition, in this paper, the problem of preserving the privacy of the systems is cast as a concrete optimization problem that balances the need for keeping the privacy with that of the maintaining the performance. This provides a different approach to that of differential privacy in which constraints on the performance degradation cannot be enforced directly to optimally balance between privacy and performance. Finally, note that the problem of releasing the dynamical model of a system under privacy constraints was considered in . In this paper, we take a different approach, i.e., we do not release the model of the system. We want to ensure that inferring an exact model relating inputs and outputs is made difficult. The rest of the paper is organized as follows. The design of optimal additive input and output noise to hinder system identification is studied in Section [2](#sec:optimalnoise){reference-type="ref" reference="sec:optimalnoise"}. Section [3](#sec:diffprivacy){reference-type="ref" reference="sec:diffprivacy"} uses the differential privacy for constructing additive output noises. A numerical example is provided in Section [4](#sec:example){reference-type="ref" reference="sec:example"}. Some concluding remarks are presented in Section [5](#sec:conc){reference-type="ref" reference="sec:conc"}.
# Optimal Additive Noise {#sec:optimalnoise}
Here, we investigate the use of additive noise to preserve the privacy of the model information assuming that the eavesdropper uses the best linear unbiased estimate. These results are subsequently generalized (to the case where the model of the eavesdropper is not known) when using the differential privacy framework.
## Problem Formulation {#sec:problem}
In this paper, The system is described by the following equation \[y_t = H(q^{-1})r_t + e_t,\] where \(H(q^{-1})\) represents the transfer function of the system, which is driven by the reference input \(r_t\). The output \(y_t\) is corrupted by additive white Gaussian noise with variance \(\sigma^2\), which is represented by \(e_t\). Assume that \(H(q^{-1})\) can be well-represented by a finite-impulse response (FIR) system of order \(n_h\), i.e., \(H(q^{-1}) = \sum_{k=0}^{n_h-1} h_k q^{-k}\). Hence, the dynamics of the system is completely characterized by the vector of coefficients \(h := [h_0\,\ldots\,h_{n_h-1}]^\top\). Assume that an adversary is interested in inferring on the process relating \(r_t\) to \(y_t\) by attempting to estimate \(h\) from a set of \(N\) input/output measurements \(\{r_t,\,y_t\}_{t=1}^N\). To complicate the identification process, an additional component (which is not accessible to the adversary) can be added to the input or to the output of the system to lower the identification accuracy. Let \(w_t\) capture such an additional component, which changes the model of the system as \[y_t = H(q^{-1})r_t + e_t+w_t.\] This term can capture both the additive input and output noise as discussed, in detail, in what follows.
This assumption is rather conservative. When using the differential privacy framework in the next section, we can avoid such assumptions. Considering a FIR model for the system and in light of Assumption [\[assume:1\]](#assume:1){reference-type="ref" reference="assume:1"}, the best linear unbiased estimate (BLUE) of \(h\) from perspective of the malicious entity is given by the standard least-squares estimate . Let us introduce the vectors \(y := [y_1 \,\ldots\, y_N]^\top\), \(e:= [e_1 \,\ldots\, e_N]^\top\), and \(w:= [w_1 \,\ldots\, w_N]^\top\). Assuming that the system is at rest prior to the data collection (i.e., \(r_t=0\) for all \(t\leq 0\)) and defining the matrix \[R := \begin{bmatrix} r_1 & 0 & 0 & \ldots & 0 \\ r_2 & r_1 & 0 & \ldots & 0 \\ \vdots & \vdots & \vdots &\ddots & \vdots \\ r_{n_h} & r_{n_h-1} & r_{n_h-2} &\ldots & r_1 \\ \vdots & \vdots & \vdots &\ddots & \vdots \\ r_{N} & r_{N-1} & \ldots &\ldots & r_{N-n_h+1} \end{bmatrix}\hspace{-.04in},\] it is evident that \(y = Rh + w + e.\) The least-squares estimate of \(h\) is then given by \[\label{eq:LS_est} \hat h = (R^\top R)^{-1}R^\top y.\] Note that this estimator is not the true BLUE, which would require the knowledge of the second order statistics of \(w_t\). However, it is the best that the malicious entity can do without the knowledge that \(w_t\) exists. This estimator is still unbiased because \({\mathbb E} \{ \hat h \} = {\mathbb E} \{(R^\top R)^{-1}R^\top (Rh + w + e) \} = h + (R^\top R)^{-1}R^\top {\mathbb E} \{w + e \} = h\). Then, a measure of the accuracy of the estimation of the impulse response is the covariance matrix of \(\hat h\) , namely \[P_h := {\mathbb E} \{(\hat h-h)(\hat h-h)^\top \}.\] The additional input \(w_t\) determines the quality of the estimated system \(\hat h\) by entering into the expression of the parameter covariance matrix \(P_h\). Intuitively, the higher the power of \(w_t\), the higher \(P_h\) (and thus the lower the identification accuracy). On the other hand, \(w_t\) has an undesired effect on the output power. Therefore, the additive noise is designed to increase the total variance of \(\hat h\) (expressed through the trace of \(P_h\)) while keeping low the contribution of \(w_t\) to the variance of \(y_t\). .
Here, the additive noise is designed for a given sequence of inputs captured by \(r\). This might not be generally feasible as, when dealing with causal systems, the additive noise should be designed and employed prior to receiving the entire sequence of inputs. This design methodology is however very useful to make the identification difficult for a given input, such as those in optimal experimental design in the model identification literature . Alternatively, a distribution for the input signal can be considered.
In this case, the identification error \(P_h\) which is used as a measure of privacy should be replaced with \(\mathbb{E}\{P_h\}\) with the expectation being taken over random variables \(r\) and \(N\). This allows us to generalize the problem of the interest as follows.
In this paper, two families of additive noise are considered, namely, additive output noise and additive input noise. In the remainder of this section, these two families are described.
### Additive Output Noise
The additive noise \(w_t\) is modelled by a zero-mean moving-average (MA) stochastic process of the form \[w_t = L(q^{-1}) v_t,\] where \(v_t\) is a sequence of i.i.d. zero-mean noise () of unit variance and \(L(q^{-1}):=\sum_{k=0}^{n_l}l_kq^{-k}\) is a FIR filter of prescribed order \(n_l\). . The additive noise \(w:=[w_1\,\ldots\, w_N]^\top\) can be expressed as \(w=Lv\), where \(v := [v_{-n_l+2} \,\ldots v_0 \, v_1 \,\ldots\, v_N]^\top\) and \[\label{eq:def_L} L:= \begin{bmatrix} l_{n_l-1} & \ldots & l_0 & 0 & 0 & \ldots & 0 \\ 0 & l_{n_l-1} & \ldots & l_0 & 0 & \ldots & 0\\ \vdots & \ddots & \ddots & \ddots & \ddots & \ddots & \vdots \\ 0 & \ldots & 0 & l_{n_l-1} & \ldots & l_0 & 0 \\ 0 & 0 & \ldots & 0 & l_{n_l-1} & \ldots & l_0 \end{bmatrix}.\] The identification error , in this case, is \[\begin{aligned}
\label{eq:ls_variance:outputnoise} P_h & = (R^\top R)^{-1}R^\top \mbox{\rm Var} [w+e] R (R^\top R)^{-1} \nonumber \\ & = (R^\top R)^{-1}R^\top(L L^\top + \sigma^2I_N)R (R^\top R)^{-1}. \end{aligned}\] Further, the output variance can be determined by \[\begin{aligned}
\lambda_y := {\mathbb E} \{y_t^2|r_t = 0 \ = {\mathbb E} \{(w_t + e_t)^2 \ = \|l\|^2 + \sigma^2, \end{aligned}\] where \(l=[l_0 \,\ldots\, l_{n_l-1}]^\top\).
### Additive Input Noise
In this case, the additive input noise is denoted by \(x_t\) and is modeled by a zero-mean MA stochastic process of the form \[x_t = L(q^{-1}) v_t,\] where, similarly, \(v_t\) is a sequence of i.i.d. zero-mean noise of unit variance and \(L(q^{-1})\) is a FIR filter of prescribed order \(n_l\) determining the autocorrelation of \(x_t\). Then, the new system is described by \[\begin{aligned}
y_t &= H(q^{-1})(r_t + x_t) + e_t \nonumber \\ &= H(q^{-1})(r_t + L(q^{-1}) v_t) + e_t. \end{aligned}\] The additive noise \(w_t\), in this case, is the contribution of \(x_t\) to the output, i.e., \(w_t=H(q^{-1})L(q^{-1}) v_t\). Define \[F(q^{-1}) := H(q^{-1})L(q^{-1}),\] which can be expressed as \[F(q^{-1}) = \sum_{k=0}^{n_f-1} f_k q^{-k}, \; n_f\hspace{-.03in} =\hspace{-.03in} n_h\hspace{-.03in}+\hspace{-.03in}n_l\hspace{-.03in}-\hspace{-.03in}1.\] Note that \(x\hspace{-.03in}:=\hspace{-.03in}[x_1\ldots x_N]^\top\) can be expressed as \(x\hspace{-.02in}=\hspace{-.02in}Fv\) with \(v := [v_{-n_f+2}\ldots v_0 v_1 \,\ldots v_N]^\top\) and \(F\) is defined similarly to \(L\) in [\[eq:def_L\]](#eq:def_L){reference-type="eqref" reference="eq:def_L"}. The identification error becomes \[\begin{aligned}
\label{eq:ls_variance} P_h & = (R^\top R)^{-1}R^\top \mbox{\rm Var} [w+e] R (R^\top R)^{-1} \nonumber \\ & = (R^\top R)^{-1}R^\top(F F^\top + \sigma^2I_N)R (R^\top R)^{-1}. \end{aligned}\] Finally, it can be shown that \(\lambda_y = \|f\|^2 + \sigma^2\), where \(f=[f_0 \,\ldots\, f_{n_f-1}]^\top\).
## Deterministic Input {#sec:determ}
This part is dedicated to solving Problem [\[prob:det\]](#prob:det){reference-type="ref" reference="prob:det"}. The results are first presented for the output noise case.
### Additive Output Noise
For additive output noise, Problem [\[prob:det\]](#prob:det){reference-type="ref" reference="prob:det"} can be rewritten as
where \(\gamma_1\) denotes the maximum tolerated output variance. Define the performance degradation ratio \[\begin{aligned}
\rho :=&\frac{{\mathbb E} \{y_t^2|r_t = 0 \}}{{\mathbb E} \{y_t^2|r_t = 0,w_t=0 \}} =\frac{\lambda_y}{\sigma^2}. \end{aligned}\] If the goal of the designer is to keep the performance degradation ratio below \(\epsilon\), the constant \(\gamma_1\) can be selected to be smaller than \(\sigma^2 \epsilon\). Defining \(M:= Q_l^\top ( I_{N+n_l-1} \otimes E) Q_l\) and noting that the term \(c\) is independent of \(l\) (and thus can be discarded from the optimization problem), we transform [\[eq:criterion_1:output\]](#eq:criterion_1:output){reference-type="eqref" reference="eq:criterion_1:output"} into
The following result can be immediately proved.
\[\begin{aligned}
\label{eq:criterion_3:output} \argmin_{l \in \mathbb{R}^{n_l}} \left(\mbox{\rm tr}(P_h)\right)^{-1} + \gamma_2 \lambda_y \,, \end{aligned}\] where \(\gamma_2\) determines weight on the performance versus the privacy. This formulation is useful when the constraint on the performance is not hard (i.e., the degradation does not need to be maintained under a given level but large output variations are not pleasant). This problem is rewritten as \[\begin{aligned}
\label{eq:criterion_4:output} \argmin_{l \in \mathbb{R}^{n_l}} (l^\top M l + c)^{-1} + \gamma_2 \|l\|^2 \,, \end{aligned}\] where \(c\) is defined in [\[eq:E_c\_def\]](#eq:E_c_def){reference-type="eqref" reference="eq:E_c_def"}.
### Additive Input Noise
Similarly, Problem [\[prob:det\]](#prob:det){reference-type="ref" reference="prob:det"} can be expressed as
Using the same line of reasoning as in Lemma [\[th:dummy_lemma\]](#th:dummy_lemma){reference-type="ref" reference="th:dummy_lemma"}, we introduce the following instrumental result.
Now, note that the coefficients of the filter \(L(q^{-1})\) and filter \(F(q^{-1})=H(q^{-1})L(q^{-1})\) are related according to \[\label{eqn:f_h_l} f = Hl,\] where \(H \in \mathbb{R}^{n_f \times n_l}\) is a Toeplitz matrix formed by the coefficients of \(h\). Substituting [\[eqn:f_h\_l\]](#eqn:f_h_l){reference-type="eqref" reference="eqn:f_h_l"} in [\[eqn:error_to_f\]](#eqn:error_to_f){reference-type="eqref" reference="eqn:error_to_f"} gives \(\mbox{\rm tr}(P_h) = l^\top H^\top Q_f^\top ( I_{N+n_f-1} \otimes E ) Q_f H l+c.\) Therefore, the optimization problem in [\[eq:criterion_1:input\]](#eq:criterion_1:input){reference-type="eqref" reference="eq:criterion_1:input"} can be transformed into
where \(M'=H^\top Q_f^\top ( I_{N+n_f-1} \otimes E ) Q_f H\). The following result can be immediately proved.
The condition \(H^\top H>0\) is satisfied so long as \(H\) has full column rank. This is guaranteed if \(h_{n_h}\neq 0\), i.e., no fewer than \(n_h\) parameters are required for describing filter \(H(q^{-1})\).
## Random Inputs {#sec:random}
The problem of designing an additive output noise is only considered in this section. The results can be easily extended to the design of input noises following the same line of reasoning. Problem [\[prob:sto\]](#prob:sto){reference-type="ref" reference="prob:sto"} can be cast as
Note that \(\mbox{\rm tr}(P_h) =\mathbb{E}\{c(r,N)\}+ l^\top \mathbb{E}\{Q_l(N)^\top ( I_{N+n_f-1} \otimes E(r,N) ) Q_l(N) \}l\). Although having the same definition, \(Q_l(N)\), \(E(r,N)\), \(c(r,N)\) are used instead of \(Q_l\), \(E\), and \(c\) to emphasize they are functions of random variables \(N\) and \(r\). Define \(M'':=\mathbb{E}\{Q_l(N)^\top ( I_{N+n_f-1} \otimes E(r,N) ) Q_l(N) \}\). The optimization problem in [\[eqn:optim:sto\]](#eqn:optim:sto){reference-type="eqref" reference="eqn:optim:sto"} can be rewritten as
Unfortunately, calculating \(M''\) in an explicit from as a function of the distributions of \(N\) and \(r\) is generally difficult. The following remark provides a numerical algorithm for constructing an approximation of this matrix.
# Relationship to Differential Privacy {#sec:diffprivacy}
Throughout this section, the design of an additive output noise is only considered. The results for the additive input noise can be constructed similarly. Furthermore, \(h\) is assumed to belong to a compact set \(\mathcal{H}\subseteq \mathbb{R}^{n_h}\).
Note that a random variable \(w\) is said to follow the Laplace distribution with mean \(\mu\) and (scaling) parameter \(b>0\) if \(\mathbb{P}\{w\in\mathcal{W}\}=\int_{w\in\mathcal{W}} (2b)^{-1} \exp(-|w-\mu|/b)\mathrm{d}w\) for all Lebesgue-measurable sets \(\mathcal{W}\subseteq \mathbb{R}\).
Note that \(\sup_{h,h'\in\mathcal{H}:\|h-h'\|_0\leq 1} \|Rh-Rh'\|_1\) exists and is finite because \(\mathcal{H}\) is assumed to be a compact set.
Combination of Theorems [\[tho:diffprivacy\]](#tho:diffprivacy){reference-type="ref" reference="tho:diffprivacy"} and [\[tho:usefulness\]](#tho:usefulness){reference-type="ref" reference="tho:usefulness"} illustrates the trade-off between preserving privacy and closed-loop performance because as \(\epsilon\) tends to zero (to achieve a higher level of privacy), the performance degrades (i.e., \(\lambda_y\) goes to infinity).
Proposition [\[prop:b\]](#prop:b){reference-type="ref" reference="prop:b"} illustrates that the parameter of the Laplace noise \(b\) should be increased upon admitting larger input sequences. This is because, with larger \(N\), there are more data to extract the system parameters and, thus, the employed mechanism needs to be more conservative to avoid leaking the private information. Some relaxations of the differential privacy, e.g., \((\epsilon,\delta)\)-differential privacy, that lend themselves to using a Gaussian noise, e.g., . Let for any \(\epsilon\) and \(\delta\) define \(\kappa(\epsilon,\delta)=(\mathcal{Q}^{-1}(\delta)+\sqrt{\mathcal{Q}^{-1}(\delta)^2+2\epsilon})/2\) with \(\mathcal{Q}^{-1}\) denoting the inverse of \(\mathcal{Q}:x\mapsto \int_{x}^\infty 1/\sqrt{2\pi}\exp(-u^2/2)\mathrm{d}u\).
# Numerical Examples {#sec:example}
Consider the discrete-time system \(y_t=G(q^{-1})r_t+e_t,\) where \(G(q^{-1})=(q^{-1}-0.2q^{-2})/( 1-0.9 q^{-1} + 0.17q^{-2}).\) Clearly, \(G(q^{-1})\) is not a FIR system. This system can be approximated by the FIR filter \(H(q^{-1})= q^{-1}+0.7q^{-2}+0.46q^{-3}+0.295q^{-4} +0.1873q^{-5}+0.1184q^{-6}+0.0747q^{-7} +0.0471q^{-8}+0.0297q^{-9}.\) The quality of the approximation is \(\|H(q^{-1})-G(q^{-1})\|=0.0507\).
### Random inputs {#random-inputs}
Assume that the malicious entity injects a sequence of i.i.d. zero-mean unit-variance Gaussian variables of length \(N\) chosen with equal probability from \(\{10,\dots,20\}\). The approach of Subsection [2.3](#sec:random){reference-type="ref" reference="sec:random"} is considered for constructing an optimal additive output noise with \(n_l=5\). In this example, \(M''\) is approximated using the method of Remark [\[remark:1\]](#remark:1){reference-type="ref" reference="remark:1"} with \(\theta=100\) and \(\vartheta=1000\). Set \(\sigma^2=0.1\) and \(\gamma_1=0.2\). Therefore, the performance degradation ratio is upper-bounded as \(\rho\leq 2\) (indeed the upper bound is tight due to the nature of the optimal solution). The optimal additive input noise, in this case, is driven by the FIR filter \(L(q^{-1})=0.1450+0.0799q^{-1}+0.2125q^{-2} +0.0799q^{-3}+0.1450q^{-4}\). Using the Monte Carlo simulation, it can be shown that \(\mbox{\rm tr}(\mathbb{E}\{P_h\})/\mbox{\rm tr}(\mathbb{E}\{P_h|w_t=0\})\approx 1.9639.\) Therefore, the system identification error has been approximately doubled at the expense of doubling the output variance. From Theorem [\[tho:random\]](#tho:random){reference-type="ref" reference="tho:random"}, it can be inferred that \(\mbox{\rm tr}(\mathbb{E}\{P_h\})/\mbox{\rm tr}(\mathbb{E}\{P_h|w_t=0\})=1+(\eta^{*\top} M''\eta^*)/\mathbb{E}\{c(r,N)\}(\gamma_1-\sigma^2).\)
# Conclusions {#sec:conc}
Adding input and output noises for increasing the model identification error was considered. Optimal filters for constructing additive coloured noises were designed to maximize the identification error while maintaining the closed-performance degradation below a threshold. Differential privacy was also explored for designing output noises that preserve the privacy of the model.
[^1]: i.i.d. stands for independently and identically distributed. | {'timestamp': '2017-06-07T02:02:59', 'yymm': '1706', 'arxiv_id': '1706.01587', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01587'} |
# Introduction {#sec:intro}
A pure superfluid in the zero temperature limit has no viscosity. The superfluid is formed by atoms which have condensed into the ground state. The condensate atoms behave collectively and are described by a coherent macroscopic wavefunction. In liquid \(^3\)He and dilute Fermi gases, the condensate is formed by pairs of atoms known as Cooper pairs. For most superfluids, including \(^3\)He-B, the superfluid velocity **v** is proportional to the gradient of the phase of the wavefunction, \({\mbox{\boldmath \)\nabla\(}}\theta\). Thus the superfluid flow is irrotational \({\mbox{\boldmath \)\nabla\(}}\times{\bf {v}} ={\bf 0}\) (more complex superfluids such as \(^3\)He-A are not considered here). Superfluids can however support line defects on which \({\mbox{\boldmath \)\nabla\(}}\times{\bf {v}}\) is singular. These defects are quantum vortices. The phase \(\theta\) of the wavefunction changes by \(2\pi\) around the vortex core and gives rise to a circulating flow. Each vortex carries a single quantum of circulation \(\kappa = 2\pi\hbar/m\) where \(m\) is the mass of the constituent atoms or atom-pairs (for superfluid \(^3\)He-B, \(\kappa = \pi\hbar/2m_3=0.662\times10^{-7}\,{\rm m^2/s}\), where \(m_3\) is the mass of a \(^3\)He atom). At low temperatures, quantum vortices move with the local superfluid velocity. Instabilities and reconnections of quantized vortices result in the formation of a vortex tangle which displays complex dynamics, as each vortex moves according to the collective velocity field of all other vortices. This complex, disordered flow is known as quantum turbulence. A property that characterizes the intensity of quantum turbulence is the vortex line density, \(L\) (\({\rm m}^{-2}\)), that is, the total length of vortex lines per unit volume. At higher temperatures the thermal excitations form a normal fluid component which may couple to the superfluid via mutual friction. The resulting behavior is quite complex, depending on whether the normal fluid also becomes turbulent. Here we only consider the low temperature regime in \(^3\)He-B, \(T\lesssim0.3T_c\), where \(T_c\approx0.9\,{\rm mK}\), is the critical temperature of superfluid transition at zero pressure. In this temperature regime, thermal excitations no longer form the normal fluid, but a few remaining excitations, whose mean free paths greatly exceed the typical size of the experimental cell, form a ballistic gas which has no influence on the vortex dynamics. Even in this low temperature case, the properties of quantum turbulence are not yet fully understood; besides, it is interesting to compare the properties of pure quantum turbulence with those of classical turbulence. Understanding their relationship may eventually lead to a better understanding of turbulence in general. During the last two decades there has been significant progress in the experimental studies of quantum turbulence. Experimental techniques are being developed to measure properties of quantum turbulence at low temperatures, both in \(^4\)He (see e.g. Refs. ) and in \(^3\)He-B. The most promising technique which is being developed for superfluid \(^3\)He-B utilizes the Andreev reflection of quasiparticle excitations from superfluid flow . (This technique was first developed to study quantum turbulence in stationary samples , and has since been extended to rotating samples .) Andreev reflection arises in a Fermi superfluid as follows. The energy-momentum dispersion curve \(E({\bf {p}})\) for excitations has a minimum at the Fermi momentum, \({\mbox{\)p_F\(}}\), corresponding to the Cooper pair binding energy \(\Delta\). Quasiparticle excitations have \(|{\bf {p}}| > {\mbox{\)p_F\(}}\) whereas quasiholes have \(|{\bf {p}}|< {\mbox{\)p_F\(}}\). On moving from one side of the minimum to the other, the excitation's group velocity reverses: quasiholes and quasiparticles with similar momenta travel in opposite directions. In the reference frame of a superfluid moving with velocity \({\bf {v}}\), the dispersion curve tilts to become \(E({\bf {p}})+{\bf {p}}\cdot{\bf {v}}\). Quasiparticles which move into a region where the superfluid is flowing along their momentum direction will experience a potential barrier. Quasiparticles with insufficient energy are reflected as quasiholes. The momentum transfer is very small so there is very little force exerted whilst the outgoing quasiholes almost exactly retrace the path of the incoming quasiparticles . Andreev reflection therefore offers an ideal passive probe for observing vortices at very low temperatures. Numerical simulations of Andreev reflection for two-dimensional vortex configurations have shown that the total reflection from a configuration of vortices may be significantly less than the sum of reflections from individual vortices . This phenomenon is called "partial screening". The Andreev scattering cross-sections of three-dimensional isolated vortex rings and the role of screening effects from various configurations of vortex rings in \(^3\)He-B were studied in Ref. . In this work we analyze numerically the Andreev reflection of quasiparticles by three-dimensional vortex tangles in \(^3\)He-B. We use moderate tangle densities which are comparable to those typically observed in the experiments. The plan of the paper is as follows. In Sec. [2](#sec:theornumer){reference-type="ref" reference="sec:theornumer"} we introduce the basic equations governing the dynamics of the vortex tangle and the associated flow field, the equations governing the propagation of thermal quasiparticle excitations, and the method for the calculation of the Andreev-reflected fraction of quasiparticle excitations incident on the vortex tangle. In Sec. [\[sec:vortgeneration\]](#sec:vortgeneration){reference-type="ref" reference="sec:vortgeneration"} we describe our method for the numerical generation of vortex tangles and discuss their properties. In Sec. [\[sec:quasiclassical\]](#sec:quasiclassical){reference-type="ref" reference="sec:quasiclassical"} we present our results for the simulation of the Andreev reflection from saturated, quasiclassical vortex tangles and analyze the dependence of the total reflection coefficient and the reflection scattering length on the vortex line density. In subsection [\[sec:exptl\]](#sec:exptl){reference-type="ref" reference="sec:exptl"} we describe the experimental techniques and the results for the experimental measurements of the Andreev reflection from quantum turbulence for comparison with our simulations. In Sec. [\[sec:ultraquantum\]](#sec:ultraquantum){reference-type="ref" reference="sec:ultraquantum"} we compare these results with those obtained for the Andreev reflection from unstructured, 'ultraquantum' turbulence. In Sec. [\[sec:screening\]](#sec:screening){reference-type="ref" reference="sec:screening"} we discuss the mechanisms of screening in the Andreev reflection from vortex tangles. In Sec. [3](#sec:fluctuations){reference-type="ref" reference="sec:fluctuations"} we present our numerical results for the spectral properties of the Andreev reflection and of the vortex line density, analyze the correlation between these two statistical properties and compare them with the statistical properties of the experimentally-observed Andreev-retroreflected signal. We discuss our findings and their interpretation in Sec. [4](#sec:interpretations){reference-type="ref" reference="sec:interpretations"}.
# Dynamics of the vortex tangle and propagation of thermal quasiparticle excitations: governing equations and numerical methods {#sec:theornumer}
Since our spatial resolution scale is much larger than the coherence length, we will model the vortex filaments as infinitesimally thin vortex lines. At low temperatures, \(T\lesssim0.3T_c\) such that mutual friction can be ignored, the superflow field \({\bf {v}}({\bf {r}},\,t)\) and the dynamics of the vortex tangle are governed by the coupled equations \[{\bf {v}}({\bf {s}},\,t)=-\frac{\kappa}{4\pi}\oint\limits_{\cal L} \frac{({\bf {s}}-{\bf {r}})}{\vert{\bf {s}}-{\bf {r}}\vert^3}\times d{\bf {r}}\,, \quad \frac{d{\bf {s}}}{dt}={\bf {v}}({\bf {s}},\,t)\,, \label{eq:biot_savart}\] where the Biot-Savart line integral extends over the entire vortex configuration \(\cal L\), with \({\bf {s}}={\bf {s}}(t)\) identifying a point on a vortex line. In the first equation, the singularity at \({\bf {r}}={\bf {s}}\) is removed in a standard way . The second, kinematic equation states that the vortex line moves with the local superfluid velocity. This is valid at low temperatures where mutual friction is absent . Our numerical simulation of Eqs. ([\[eq:biot_savart\]](#eq:biot_savart){reference-type="ref" reference="eq:biot_savart"}) is made in a periodic cubic box for vortex tangles of line densities up to \(L \sim 10^8\,{\rm m^{-2}}\), see Sec. [\[sec:vortgeneration\]](#sec:vortgeneration){reference-type="ref" reference="sec:vortgeneration"} for detail. In Ref. , based on the approach developed by Greaves and Leggett , we derived the following semiclassical Hamiltonian equations governing the propagation of thermal quasiparticle excitations in \(^3\)He-B: \[\frac{d{\bf {r}}}{dt}=\frac{{\mbox{\(\epsilon_p\)}}}{\sqrt{{\mbox{\(\epsilon_p\)}}^2+\Delta^2}}\frac{{\bf {p}}}{m^*}+{\bf {v}}, \quad \frac{d{\bf {p}}}{dt}=-\frac{\partial}{\partial{\bf {r}}}({\bf {p}}\cdot{\bf {v}})\,, \label{eq:motion}\] where \({\bf {r}}(t)\) represents the position of a quasiparticle excitation taken as a compact object, \({\bf {p}}(t)\) is the momentum of excitation, \(\epsilon_p=p^2/(2m^*)-\epsilon_F\) is the "kinetic" energy of a thermal excitation relative to the Fermi energy, \(m^*\approx3.01m_3\) is the effective mass of an excitation, with \(m_3\) being the bare mass of a \(^3\)He atom, and \(\Delta\) is the superfluid energy gap. Equations ([\[eq:motion\]](#eq:motion){reference-type="ref" reference="eq:motion"}) were derived by considering the propagation of excitations at distances from the vortex core large compared to the coherence length (at zero temperature and pressure the coherence length is \(In principle, interactions of thermal excitations with the vortex tangle can be analyzed by simultaneously solving Eqs.~(\ref{eq:motion}), governing the propagation of quasiparticles, and Eqs.~(\ref{eq:biot_savart}), governing the dynamics and the superflow field of the tangle. This is exactly how some of the authors of this paper earlier analyzed the Andreev reflection from two-dimensional vortex configurations~ and three-dimensional quantized vortex rings~. However, Eqs.~(\ref{eq:motion}) are stiff, and solving them simultaneously with Eqs.~(\ref{eq:biot_savart}) for saturated three-dimensional vortex tangles requires considerable computing time and resources. For tangles of moderate line densities (up to\)L \~10\^8 m\^-2\() considered in this work, an alternative, much simpler method for evaluating the flux of quasiparticles Andreev-reflected by the tangle can be developed based on the properties~ of solution of Eqs.~(\ref{eq:motion}) and on the arguments~ formulated below. The alternative method makes use of the facts~; first, that at moderate tangle densities, the Andreev reflected excitation practically retraces the trajectory of the incident quasiparticle, and secondly that the time scale of propagation of the quasiparticle excitation within the computational box (of 1~mm size in our simulations) is much shorter than the characteristic timescale,\)\_f\~1/(L)\(of the fluid motion in the tangle. In analyzing the quasiparticle trajectories, these results allow us to assume a `frozen' configuration of vortices and the associated flow field. Assuming that the statistically uniform and isotropic vortex tangle occupies the cubic computational box of size\)D\(, consider the propagation of a flux of quasiparticles at normal incidence on one side of the box in (say) the\)x\(-direction. Also assume that the quasiparticle flux is uniformly distributed over the\)(y, z)\(plane and covers the full cross section of the computational box. Ignoring angular factors, the incident flux of quasiparticles, as a function of the position\)(y, z)\(, can be written as~ \begin{equation} \langle nv_g\rangle^i_{(y,z)}=\int^{\infty}_{\Delta}g(E)f(E)v_g(E)dE\,, \label{eq:Eflux} \end{equation} where\)g(E)\(is the density of states, and\)f(E)\(is the Fermi distribution function which we can approximate at low temperatures to the Boltzmann distribution,\)f(E) = (-E/k_BT)\(. The average energy of a quasiparticle in this flux is\)E= + k_BT\(. Since the typical quasiparticle group velocity is large compared to a typical superfluid flow velocity\)v\(, the integral is simplified by noting that\)g(E)v_g(E) = g(p_F)\(is the constant density of momentum states at the Fermi surface. Equation~(\ref{eq:Eflux}) therefore reduces to \begin{align} \begin{split} \langle nv_g\rangle^i_{y,z}&=g({\mbox{\)p_F\(}})\int^{\infty}_{\Delta}\exp(-E/k_BT)\,dE \\ &= g(p_F) k_BT\exp(-\Delta/k_BT)\. \end{split} \label{eq:Exfluxsimp} \end{align} In the presence of a tangle, and hence regions of varying superfluid velocity\)**v**\(, a quasiparticle moving into a region where the superfluid is flowing along their momentum direction experiences a force\)=-(**p**)\(, which decreases its momentum towards\)p_F\(and reduces its group velocity. If the flow is sufficiently large the quasiparticle is pushed around the minimum, becoming a quasihole with a reversed group velocity. Consequently, the flux of quasiparticles transmitted through a tangle,\)nv_g\^t\(is calculated by the integral~(\ref{eq:Exfluxsimp}) in which the lower limit is replaced by\)+(**p**)+\(p_F\) v_x\^\(, where\)v_x\^\(is the maximum value of the\)x\(-component of superfluid velocity, found from the solution of Eqs.~(\ref{eq:biot_savart}), along the quasiparticle's rectilinear trajectory [that is, for fixed\)(y, z)\(]: \begin{align} \begin{split} \langle nv_g\rangle^t_{y,z} = g(p_F)\int^{\infty}_{\Delta + p_Fv_x^{\max}}\exp(-E/k_BT)\,dE \\ = g(p_F)k_BT\exp[-(\Delta + p_Fv_x^{\max})/k_BT]\. \end{split} \label{eq:Exfluxsimp2} \end{align} For the position\)(y, z)\(on the face of computational box, the Andreev reflected fraction of quasiparticles incident on a tangle along the\)x\(-direction is, therefore, \begin{equation} f_{y,z} = 1-\frac{\langle nv_g\rangle^t_{y,z}}{\langle nv_g\rangle^i_{y,z}} = 1-\exp\left[-\frac{p_Fv_x^{\max}}{k_BT}\right]\. \label{eq:ExAndsimp} \end{equation} The total Andreev reflection\)f_x\(is the average of the Andreev reflections for all positions of the\)(y, z)\(plane. The same calculation is repeated for the flux of quasiholes. The results are then combined to yield the reflection coefficient for the full beam of thermal excitations. In our simulation, the initial positions of quasiparticle excitations on the\)(y, z)\(plane were uniformly distributed with spatial resolution\) m\^2\(. More details of numerical simulation will be given below in Sec.~\ref{sec:vortgeneration}. In the following sections, to directly compare our simulations with experimental observations, we fix the temperature at\)T=0.15T_c\(and use parameters appropriate to\)\^3\(He-B at low pressure: Fermi energy\[\epsilon_F\)=2.27\^-23 J\(, Fermi momentum\]p_F\) = 8.28 \^-25 kg m s\^-1\(, superfluid energy gap\)\_0=1.76k_BT_c\(, and quasiparticle effective mass\)m\^\*m_3=1.51 \^-26 kg\(. \section{Numerical generation and properties of vortex tangles} \label{sec:vortgeneration} Motivated by experimental studies~, we numerically simulate the evolution of a vortex tangle driven by vortex loop injection. Two rings of radius\)R_i=240 \(are injected at opposite corners of the numerical domain (see Fig.~\ref{fig:transient}(A), where the two injected vortex rings, marked in red, can be seen entering from the left and right sides of the box repeatedly with a frequency\)f_i=10 Hz\(. The vortex loops injected into the simulation box collide and reconnect to rapidly generate a vortex tangle. The time evolution of the vortex tangle and the superfluid velocity\)**v**(**r**, t)\(were simulated by means of the vortex filament method, that is by solving Eqs.~(\ref{eq:biot_savart}), in a cubic box of size\)D = 1 mm\(with periodic boundary conditions~. (As can be seen from Eq.~(\ref{eq:ExAndsimp}), the accuracy of simulation of the superflow field is crucially important for the calculation of the Andreev reflection.) To ensure good isotropy, the plane of injection of the loops is switched at both corners with a frequency of\)f_s=3.3 Hz\(. We integrate Eqs.~(\ref{eq:biot_savart}) for a period of 380~s. For dense vortex tangles, we use the tree-method with a critical opening angle of 0.4 rad~. For small line densities, of the order of few\)mm\^-2\(or less, since the tree-method loses accuracy, calculations were performed instead by the direct evaluation of the Biot-Savart integrals. The evolution of the vortex line density for the first 40~s of the simulation is shown in Fig.~\ref{fig:transient}. After an initial transient, lasting for about 5~s, the energy injected into the box is balanced by the numerical dissipation, which arises from the spatial resolution (\[6~\]m) of the numerical routine (small scale structures such as small vortex loops and high frequency Kelvin waves of wavelength shorter than this spatial resolution are therefore lost), and the tangle saturates to reach the equilibrium state. In this statistically steady state the vortex line density,\)L(t)\(, fluctuates around its equilibrium value\)L= 9.7 \^7\(m\)\^-2\(which corresponds to the characteristic intervortex distance of\)L\^-1/2 = 102\(~\[m. A snapshot of this tangle is shown in the inset (B) of Fig.~\ref{fig:transient}. In Fig.~\ref{fig:tanglspectr} we show the energy spectrum,\)E(k)\(of this tangle (here\)k\(is the wavenumber). The energy spectrum, representing the distribution of kinetic energy over the length scales, is an important property of homogeneous isotropic turbulence. In the statistically steady state the spectrum is consistent, at moderately large scales (\)k_D \< k \< k\_\(, where\)k_D = 2/D \^3\(~m\)\^-1\(and\)k\_= 2/ \^4\(m\)\^-1\(), with the Kolmogorov\)-5/3\(spectrum shown in Fig.~\ref{fig:tanglspectr} by the red dashed line. In classical turbulence the Kolmogorov\)-5/3\(scaling is the signature of a `cascade' mechanism which transfers energy from large length scales to small length scales. The blue dotted line shows the\)k\^-1\(power spectrum expected for small scales (\)k\>k\_\() arising from the superfluid flow around individual vortex cores. Fig.~\ref{fig:tanglspectr} also shows that most of the energy of the turbulence is at the smallest wavenumbers (the largest length scales). This property is caused by partial polarization of the vortex tangle: parallel vortex lines locally come together, creating metastable vortex bundles, whose net rotation represents relatively intense large-scale flows which are responsible~ for the pile-up of energy near\)k k_D\(. \section{Andreev reflection from a quasiclassical tangle} \label{sec:quasiclassical} \subsection{Two-dimensional map of local Andreev reflections} \label{sec:2Dmap} We start with illustrating the numerical calculation, based on the method described above in Sec.~\ref{sec:theornumer}, of the Andreev reflection from an isolated, rectilinear vortex shown in Fig.~\ref{fig:ARstraight}(A). The vortex's circulating flow velocity is\)v(r)=/(2r)\(, where\)r\(is the distance from the vortex core (for theoretical analysis and analytical calculation of Andreev reflection from the rectilinear vortex see Refs.~). The velocity field around the vortex was computed using Biot-Savart integrals. To ensure the homogeneity of the velocity flow profile across the simulation volume, the actual vortex extends another four lengths above and below the cube shown in the figure. We calculate the Andreev reflection of a beam of excitations which propagate in the\)y\(-direction orthogonally to the\)(x, z)\(face of the computational cube (the black arrow shows direction of the incoming excitation). This calculation gives a measure of the Andreev reflection as a two-dimensional contour map across the full cross section of the incident quasiparticle beam. In Fig.~\ref{fig:ARstraight}, panels (B) and (C) show, respectively, the reflection of quasiparticles (\)p\>\(p_F\]) and quasiholes (\)p\<\(p_F\[) from the rectilinear, quantized vortex that has a direction of vorticity pointing up. The darker (red/yellow) regions corresponding to the higher reflectivity clearly show the ``Andreev shadows'' and are located symmetrically on the immediate right of the vortex line for quasiparticles and on the immediate left for quasiholes. Note also that for a `real' thermal beam consisting of excitations of both sorts the Andreev shadow of a single vortex should be symmetric with respect to the vortex line and can be imagined as a superposition of panels~(B) and (C). The panel (D) illustrates the total Andreev reflection from the rectilinear vortex. We use the same method of evaluating Andreev reflection for our simulated tangle. The velocity field of the vortex tangle was computed using a slightly modified tree-method with a critical opening angle of 0.2, compared to the previously published result~ where a critical opening angle of 0.4 was used. The modified tree-method algorithm continues propagation along the tree when the desired critical opening angle is reached but the local tangle polarization (defined in Refs.~) exceeds 3.8. The current simulation shows a slightly better agreement with the direct Biot-Savart calculation, but overall results obtained using the modified tree-method replicate previous conclusions. Figure~\ref{fig:2Dmap} shows a distribution of the local Andreev reflection coefficient\)f\_x,z\([calculated using Eq.~(\ref{eq:ExAndsimp})] on the\)(x, z)\(plane for a tangle with the saturated equilibrium line density\)L\^7 m\^-2\((Fig.\ref{fig:transient} inset~(B)). The darker (red/yellow) regions correspond to the higher Andreev reflection; clearly visible, very dark, thin ``threads'' correspond to a high/complete Andreev reflection which is expected in the vicinity of vortex cores. The overall distribution of dark threads corresponding to vortex cores is roughly uniform and indicates that our simulated turbulent tangle is practically homogeneous. Panel~(A) illustrates the reflection of quasiparticle excitations with\)p \> p_F\(, and (B) shows the reflection of quasiholes (\)p \< p_F\(). The shadows of large vortex rings instantaneously injected into the computational domain to sustain the statistically steady state are also visible. The individual maps of Andreev reflection for two species of excitations [quasiparticles in Fig.~\ref{fig:2Dmap}~(A) and quasiholes in Fig.~\ref{fig:2Dmap}~(B)] show the presence of large scale flow, consistent with the quasiclassical (Kolmogorov) character of the simulated vortex tangle whose energy spectrum scales as\)k\^-5/3\(in the ``inertial'' range of moderately small wavenumbers, see Fig.~\ref{fig:tanglspectr}. The extended darker regions in Fig.~\ref{fig:2Dmap}~(A) correspond to large scale flows whose\)y\(-component of velocity is directed into the page; the brighter regions show large scale flows directed out of the page. For quasiholes reflection shown in panel~(B) the meaning of darker and brighter regions is opposite to that of the panel~(A). Consequent direct calculation of the superflow patterns within the simulation volume has confirmed the presence of large scale flows in the corresponding regions. We should emphasize here that in the numerical experiment described above the \textit{direction} of the large scale flow could be determined only because the separate simulations, for the same tangle, were performed for two different species of excitations, that is for quasiparticles with\)p\>\(p_F\]and quasiholes with\)p\<\(p_F\[. In the real physical experiment the thermal beam consists of excitations of both sorts whose separation is, of course, not possible. For such a beam the Andreev reflection should be the same both for receding and approaching flows. While the presence of large scale flow may still be detected by the enhancement of the Andreev reflection, an identification of its direction would hardly be possible. A two-dimensional representation of the reflection coefficient for the \textit{full} beam of excitations Andreev-reflected by the simulated tangle is illustrated below in Fig.~\ref{fig:ARtransient} by the insets (A) and (B) for the initial and steady state of simulation, respectively. \subsection{Total Andreev reflection vs vortex line density} \label{sec:total} We now turn to the results of our calculation of the total Andreev reflection from the vortex tangle. Our aim is to find the total reflection coefficient,\)f_R\(, that is, the fraction of incident excitations Andreev-reflected by the tangle, as a function of the vortex line density. Here the total reflection coefficient is defined as\)f_R=(f_x+f_y+f_z)/3\(, where\)f_x\(,\)f_y\(, and\)f_z\(are the total reflection coefficients of the excitation beams incident, respectively, in the\)x\(,\)y\(, and\)z\(directions on the\)(y, z)\(,\)(x, z)\(, and\)(x, y)\(planes of the cubic computational domain. For the tangle numerically generated by injected rings of radius\)R_i=240 \(with the injection and switching frequencies\)f_i=10 Hz\(and\)f_i=3.3 Hz\(, respectively (see above Sec.~\ref{sec:vortgeneration} for details), we monitor simultaneously the evolution of the vortex line density (see Fig.~\ref{fig:transient}) and of the reflection coefficient calculated using the method described in Sec.~\ref{sec:theornumer} [see Eq.~(\ref{eq:ExAndsimp}) and the text that follows] during the transient period~ and for some time after the tangle has saturated at the statistically steady-state, equilibrium value of the vortex line density,\)L= 9.7 \^7\(m\)\^-2\(, see Fig.~\ref{fig:ARtransient}. After a very short initial period of less than 1~s, the vortex configuration becomes, and remains during the transient and afterwards practically homogeneous and isotropic. This enables us to re-interpret the results shown in Figs.~\ref{fig:transient} and \ref{fig:ARtransient} in the form of the Andreev-reflected fraction,\)f_R\(as a function of the vortex line density. This function~ is shown by the top panel of Fig.~\ref{fig:ARvsL}. As can be seen from Fig.~\ref{fig:ARvsL}, at small line densities,\)L\^7 m\^-2\(the reflection coefficient rises rapidly and linearly with\)L\(. During the initial stage of the generation of the tangle, the injected rings form a dilute gas and are virtually non-interacting. As the computation time advances, more rings enter the simulation box, start to interact, collide, reconnect, and eventually form a tangle which absorbs all injected rings. However, at small times, even after the tangle has been formed, the vortex line density remains relatively small so that the average distance between vortex lines is relatively large. In this case the flow around each vortex is practically the same as the flow field of a single, isolated vortex~. In a sufficiently dilute vortex configuration a screening of Andreev reflection (discussed below in Sec.~\ref{sec:screening}) can also be neglected so that the total reflection is a sum of reflections from individual vortex lines; this explains a linear growth with\)L\(of the reflection coefficient at low line densities. \subsection{Comparison with experiment} \label{sec:exptl} We compare the simulation with the experiment~ which has the configuration similar to that shown in the inset of the lower panel of Fig.~\ref{fig:ARvsL}. The vorticity is produced by the oscillating grid and surrounds the vibrating wire detectors. The tangle's flow field Andreev-reflects ambient thermal excitations arriving from ``infinity'', shielding the wires and reducing the damping. The reduction in damping on each wire (placed at 1.5\,mm, 2.4\,mm and 3.5\,mm from the grid) provides the measure of the Andreev reflection by the tangle, see Ref.~. The lower panel of Fig.~\ref{fig:ARvsL} shows the fractional reflection of quasiparticles incident on each wire. The numerical and experimental data plots in Fig.~\ref{fig:ARvsL} have similar shapes. While the vortex line density\)L\(cannot be obtained directly from the measurements, we expect the local line density of the quantum turbulence to increase steadily with increasing grid velocity. However, the onset of turbulence is rather different in the simulations and in the experiment. In the simulations, approaching injected vortex ring pairs are guaranteed to collide and form a tangle, whereas the vibrating grid emits only outward-going vortex rings, with the ring flux increasing steadily with increasing grid velocity. At low grid velocities, rings propagate ballistically with few collisions. At higher velocities, ring collisions increase giving rise to the vortex tangle. For the data of Fig.~\ref{fig:ARvsL}, this occurs when the grid velocity exceeds\)\~\(3\,mm/s. The data for lower velocities correspond to reflection from ballistic vortex rings and can be ignored for the current comparison. At higher grid velocities/tangle densities, the fraction of Andreev-reflected excitations rises at an increasing rate, finally reaching a plateau. The plateau region is prominent in the experiments when the grid reaches velocities approaching a third of the Landau critical velocity~ and may be an artifact of data analysis model that ignores a local flux of quasiparticles produced by pair-breaking. For the wire closest to the grid, the experimentally observed absolute value of the reflectivity is almost identical to that obtained from our simulation. This excellent agreement is perhaps fortuitous given that in the experiments quasiparticles travel through 1.5--2.5\,mm of turbulence to reach the wire, compared with 1\,mm in the simulation, thus larger screening might be expected. A better comparison will require high-resolution experiments to separate the effects caused by variations of tangle density from those caused by the increase of quasiparticle emission by the grid. \subsection{The Andreev scattering length} \label{sec:length} To compare the amount of Andreev scattering from vortex tangles with that of isolated vortices, it might be useful to define an Andreev scattering length. In Ref.~ the cross section,\], for Andreev scattering was defined as \begin{equation} \sigma= \frac{\dot{N}}{\langle n{\mbox{\)v_g\(}}\rangle}, \label{eq:AndScat} \end{equation} where\[is the number of excitations reflected per unit time, and\)n\(v_g\]is the number flux of incident excitations. From the cross section\[we define the Andreev scattering length,\)b\(, as \begin{equation} \sigma= b \mathcal{L}, \label{eq:AndScatlen} \end{equation} where\]is the total length of vortex line within the computational box or the experimental cell. For the simulation box with side length\)D\(,\) = LD\^3\(. From definition~(\ref{eq:AndScat}) of the cross section, the total reflectivity of a vortex configuration is\)f = /A\(, where\)A\(is the cross sectional area of the incident excitation beam. For our simulations, the Andreev scattering length is thus given by \begin{equation} b = f/(DL)\. \label{eq:AndScatlensim} \end{equation} The Andreev scattering length in the vortex tangles, obtained from Eq.~(\ref{eq:AndScatlensim}) and our numerical simulation, is plotted in Fig.~\ref{fig:s-length}. For the isolated, rectilinear quantized vortex line, the theoretical analysis and analytical calculation of the Andreev scattering length were given in Ref.~. A simpler estimate for the \textit{scale} of Andreev scattering length, used for the initial interpretation of experiments~, can be obtained as follows. Consider quasiparticle excitations incident on a straight vortex line perpendicular to the excitation flux. On one side of the vortex, an incident quasiparticle will experience a potential barrier\)p_F v = /(2m_3r)\(, where\)r\(is the impact parameter. The average thermal quasiparticle energy is\)+k_BT\(. An approximate value for the Andreev scattering length\)b\(is the impact parameter\)b_0\(at which an excitation with the average energy is Andreev reflected. Equating the potential barrier with\)k_BT\(yields \begin{equation} b_0 = \frac{p_F\hbar}{2m_3k_BT}\. \label{eq:AndImpact} \end{equation} For the parameters used in the simulations, appropriate to\)\^3\(He-B at low pressure and\)T = 150\(~\[K, estimate~(\ref{eq:AndImpact}) gives\)b_0 = 4.2\(~\]m. This value is in a very good agreement with the Andreev scattering length of the vortex tangle at the highest line densities shown in Fig.~\ref{fig:s-length}. This validates the analysis previously used in Refs.~ to interpret the experimental data for\)\^3\(He tangles. We note, however, that while yielding the correct estimate for the scale~(\ref{eq:AndImpact}) of the Andreev scattering length, this simple model might be somewhat misleading since, even for an isolated vortex, a significant amount of Andreev reflection occurs at distances larger than\)b_0\(from the vortex core, see Ref.~ for details. For dilute tangles, the Andreev scattering length is consistent with the value for large vortex rings,\)b\_rings = 10\(~\[m, inferred from the results of numerical simulations of Ref.~. It is clear that Andreev scattering from very dilute vortex tangles is dominated by reflection from individual vortices. \section{Total reflection from the unstructured vortex tangle} \label{sec:ultraquantum} In the previous section we investigated Andreev reflection from quasiclassical quantum turbulence whose energy spectrum obeys the Kolmogorov\)-5/3\(scaling law. However, there is also evidence for the existence of a second quantum turbulence regime, consisting of a tangle of randomly oriented vortex lines whose velocity fields tend to cancel each other out. This second form of turbulence, named `ultraquantum' or `Vinen' turbulence to distinguish it from the `quasiclassical' or `Kolmogorov' turbulence~, has been identified both numerically~ and experimentally (in low temperature\)\^4\(He~ and in\)\^3\(He-B~), and (only numerically) in high temperatures\)\^4\(He driven by a uniform heat current~ (thermal counterflow). Physically, ultraquantum turbulence is a state of disorder without an energy cascade. Recently, it has been argued that ultraquantum turbulence has also been observed in the thermal quench of a Bose gas~, and in small trapped atomic condensates. In the ultraquantum regime the vortex tangle is uncorrelated: There is no separation between scales (and hence no large-scale motion), and the only length scale is that of the mean intervortex distance, so that it can be expected that the total Andreev reflectivity of the ultraquantum tangle should differ from that of the quasiclassical quantum turbulence. However, we will demonstrate below that this is not the case. To compare the Andreev reflection from two different types of quantum turbulence, we should, in principle, have generated numerically the ultraquantum vortex tangle whose line density would match that of the quasiclassical tangle described above in Secs.~\ref{sec:vortgeneration} and \ref{sec:quasiclassical}. However, it turns out that a numerical generation of the ultraquantum tangle with prescribed properties presents substantial difficulties. To overcome these difficulties, we will use an alternative approach. In Refs.~ we numerically simulated vortex tangles in\)\^4\(He driven by a uniform normal fluid (to model thermally driven counterflow) and by synthetic turbulence in the normal fluid (to model mechanical driving, such as grids and propellers). Although these calculations model quantum turbulence in\)\^4\(He at elevated temperatures (\)`<!-- -->`{=html}1 KT\<T\_\(), the vortex tangle in the counterflow turbulence shares many features with the ultraquantum tangle in the zero temperature limit. Apart from small anisotropy in the streamwise vs transversal directions, the counterflow tangle is practically unstructured, with the intervortex distance being the only length scale, and the energy spectrum~ of the superfluid component, similar to that of the ultraquantum turbulence, nowhere conforms to the\)-5/3\(scaling but has a broad peak at large wavenumbers, see Fig.~\ref{fig:spectrum-ultraquantum}, followed by the same\)k\^-1\(dependence numerically observed in ultraquantum turbulence. On the other hand, the tangle simulated for the mechanically driven turbulence~ shares many features common with the quasiclassical quantum turbulence in the zero temperature regime. In particular, in both cases large-scale superflows are present, and the energy spectrum obeys Kolmogorov scaling for medium wavenumbers. Ignoring the normal fluid, the vortex tangles simulated for the thermally and mechanically driven turbulence in\)\^4\(He can be regarded as modeling, at least qualitatively, the ultraquantum and quasiclassical regimes, respectively, of quantum turbulence in the low temperature\)\^3\(He-B~. In Refs.~, both the counterflow turbulence and the vortex tangle driven by the synthetic turbulence in the normal fluid were simulated starting from an arbitrary seeding initial condition and then letting\)L\(grow and saturate to a statistically steady state. An important advantage is that in both cases the vortex line density was changing within the same interval from\)L=0\(until saturation at\)L\^8 m\^-2\(. For three realizations of the counterflow turbulence with different values of the counterflow velocity, and for three realizations of the mechanically driven turbulence with different values of the \textit{rms} normal fluid velocity (see Refs.~ for details and values of parameters), we calculated the Andreev reflection from the unstructured (``ultraquantum'') and structured (``quasiclassical'') tangles. These calculations were performed for different values of the vortex line density corresponding to different stages of the tangle's evolution from\)L=0\(to the saturated, statistically steady state. The coefficient of the total Andreev reflection was then calculated by averaging Eq.~(\ref{eq:ExAndsimp}) with the values of parameters corresponding to\)T=150 \(. The results are shown in Fig.~\ref{fig:CF-KS}. Surprisingly, for the same vortex line density the coefficients of the total Andreev reflection from the unstructured (``ultraquantum'') and the structured (``quasiclassical'') quantum turbulence appear to be practically indistinguishable, so that the measurements of the total Andreev reflection cannot be used to identify the regime of quantum turbulence~. This result is consistent with experimental observations of Andreev reflection as a function of the grid velocity, see Fig.~\ref{fig:ARvsL} and Refs.~ which show that the fully formed vortex tangle can be distinguished from the ``gas'' of individual vortex rings either by the rate of decay of the vortex line density or by the cross-correlation between different detectors. The results obtained in this and the previous sections show that the total reflection coefficient is mainly determined by the vortex line density, and that the large scale flows emerging in the quasiclassical quantum turbulence have little effect on the total Andreev reflection from the vortex tangle. This conclusion can be further strengthened by estimating the amount of scattering from large scale flows in the quasiclassical regime as follows. We assume that the largest structures generated by the simulations have a size comparable to the box length\)D\(. For classical turbulence with a Kolmogorov spectrum the characteristic velocity of the flow on a length scale\)D\(is\)v(D) C_K\^1/2\^1/3(D/2)\^1/3\(, where\)C_K \(is the Kolmogorov constant and\]is the dissipation per unit mass. For quantum turbulence, the dissipation is normally assumed~ to be given by\)= \^3 L\^2\(, where the dimensionless constant\[is inferred from experimental data~. The flow generates an energy barrier\)**p**\_F (D)\(for incident excitations. Assuming that the barrier is relatively small compared to\)k_BT\(, the corresponding total reflectivity is\)f_lp_fv(D)/(4k_BT)\(where the factor of 1/4 arises from taking an angular average of the energy barrier. The fraction of quasiparticles reflected by the large scale flow is thus estimated as \begin{equation} f_l \simeq \frac{\sqrt{C_K}}{4}\frac{p_F\kappa}{k_BT}\left(\frac{\zeta D}{2 \pi}\right)^{1/3}L^{2/3}\. \label{eq:AndFracLarge} \end{equation} This is shown by the lower black line in Fig.~\ref{fig:CF-KS}. Even at high line densities the contribution of the large scale flow to the total Andreev reflection is small (only about 15\% at the highest line density,\)L\^8 m\^-2\(; note, however, that this simple model overestimates the value of\)f_l\(). \section{Screening mechanisms in Andreev reflection from vortex tangles} \label{sec:screening} In the numerical simulations described in previous sections, as the vortex line density becomes larger than\)\~2\^7 m\^-2\(the growth of the total Andreev reflection with the vortex line density becomes more gradual, see Fig.~\ref{fig:ARvsL}. The nonlinearity of the growth of the reflection coefficient with\)L\(is due to the screening effects which become more important as the density of vortex configuration increases. We use the term `screening' to identify processes which change the overall reflectivity of the tangle for a given line density. Screening, whose mechanisms are discussed below in this section, can occur between different vortices, or between different segments of the same vortex. Analyzed in detail in Refs.~, the Andreev reflection from the flow field of an isolated, rectilinear vortex line is determined by the\)`<!-- -->`{=html}1/r\(behavior of the superfluid velocity, where\)r\(is the distance from the vortex core~. For dilute vortex configurations and tangles the total Andreev reflection is a sum of reflections from individual vortices. However, as the line density increases the total reflection no longer remains additive due to screening effects which may either reduce or enhance the Andreev reflectivity of the vortex configuration. We can identify two main mechanisms of screening. The first, which was called ``partial screening'' in earlier publications~, may significantly reduce the total reflectivity of the dense vortex configuration due to a modification of the\)`<!-- -->`{=html}1/r\(superflow field in the vicinity of the vortex line. This type of screening may occur between different vortices in a dense vortex configuration, or between different segments of the same vortex. In particular, the reflectivity may be reduced by partial cancellation of the flow produced by neighboring vortices of opposite polarity. For example, the flow field at some distance from two, sufficiently close, (nearly) antiparallel vortex lines is that of a vortex dipole,\)v\~1/r\^2\(; this, faster decay of the superflow velocity significantly reduces the total reflectivity of the vortex pair. The\)`<!-- -->`{=html}1/r\(velocity field may also be modified by the effect of large curvature of the vortex line itself. The r\^{o}le of curvature effects has been analyzed in Ref.~ for a gas of vortex rings of the same size but random positions and orientations. It was shown that the contribution to partial screening arising from curvature effects may lead to a noticeable reduction of Andreev reflection only for rings whose radii are smaller than\)R\_\* \(. It can, therefore, be expected that partial screening has another, small contribution from effects of large line curvatures arising due to high frequency Kelvin waves propagating along filaments. We may assume that in saturated vortex tangles the additional contribution to partial screening will arise due to Kelvin waves generating line curvatures,\)C=d\^2**s**/d \]\int\limits_{C^*}^\infty{\cal F}(C)\,dC\. \label{eq:KWcontrib}\[ Numerical estimates of integral ([\[eq:KWcontrib\]](#eq:KWcontrib){reference-type="ref" reference="eq:KWcontrib"}) for the considered saturated tangle of the density \(L=9.7\times10^7\,{\rm m}^{-2}\) show that the contribution of high frequency Kelvin waves to partial screening and hence the reduction of the total Andreev reflection from the tangle is negligibly small (less than 1%) . The other mechanism is that of 'fractional' screening (also called 'geometric' screening in Ref. ). This mechanism is of particular importance for dense or/and large tangles: Vortices located closer to the source of excitations screen those located at the rear. This screening mechanism will also be important in the case where, in dense tangles, vortex bundles may form producing regions of high Andreev reflectivity surrounded by those whose reflectivity is much lower. This will generate a significant 'geometric' screening since few excitations can reach vortices at the rear of a large vortex bundle, and adding a vortex to a region where there is already an intense reflection will not have a noticeable effect on the total reflectivity. Earlier a simple model of the fractional screening was proposed to infer a rough estimate for the vortex line density from the experimentally measured Andreev reflection. Assuming that all vortex lines are orthogonal to the beam of excitations, and modeling the Andreev reflection from each vortex as that from an isolated, rectilinear vortex line, the total fraction of excitations, Andreev-reflected by the tangle whose thickness in the direction of quasiparticles' beam is \(d\), was found to increase with \(d\) as \(f_R=1-\exp(-d/\lambda)\), where \(\lambda=P^{-1}\) is the lengthscale of exponential decay of the flux transmitted through the tangle and \(P=b_0L\) is the probability of reflection per unit distance within the tangle, with \(b_0\) given by Eq. ([\[eq:AndImpact\]](#eq:AndImpact){reference-type="ref" reference="eq:AndImpact"}) being the scale of Andreev scattering length by a rectilinear vortex. Then the vortex line density can be inferred from the experimentally measured value of \(f_R\) as \]L=-\gamma\ln(1-f_R)\,, \label{eq:Lvsf}\[ where, for the considered model , \(\gamma=(b_0d)^{-1}\). As noted in Ref. , formula ([\[eq:Lvsf\]](#eq:Lvsf){reference-type="ref" reference="eq:Lvsf"}) gives only rough estimates for \(L\) which are correct within a factor of 2. This is not surprising since the model ignores an influence of geometry and orientation of vortex lines on the Andreev reflection. This model can be improved by modeling the tangle as a collection of vortex rings of the same radius, \(R\), but of random orientations and positions. Andreev reflection from such vortex configurations was analyzed in more detail in our earlier work . Following Ref. , consider a thin slab of superfluid, of thickness \(\Delta x\) and cross section \(S\), but consisting now of \(N\) randomly oriented and positioned vortex rings of radius \(R\), see Fig. [\[fig:slab\]](#fig:slab){reference-type="ref" reference="fig:slab"}. The total cross section of Andreev scattering is \(\sigma_{tot}=\langle\sigma\rangle N\), where \(\langle\sigma\rangle\) is the Andreev cross section of the vortex ring, ensemble-averaged over all possible ring orientations, and \(N\) is the total number of vortex rings within the considered slab. In the case where the ring is not too small, \(R>R_*\approx2.4\,\mu{\rm m}\), it was found that \(\langle\sigma\rangle\approx2\pi K_lR\), with \(K_l=10\,\mu{\rm m}\) at temperature \(T=100\,\mu{\rm K}\). The total number of vortex rings within the slab can be calculated as \(N=\Lambda/(2\pi R)\), where \(\Lambda=LS\Delta x\) is the total line length. The probability that an incident excitation will be Andreev-reflected is \(\Delta p=\sigma_{tot}/S=\langle\sigma\rangle L\Delta x/(2\pi R)\), so that the lengthscale of decay the quasiparticle flux is \(\lambda=\Delta x/\Delta p=1/(K_lL)\). Then the line density inferred from the fraction of excitations Andreev-reflected by the tangle of thickness \(d\) is given by formula ([\[eq:Lvsf\]](#eq:Lvsf){reference-type="ref" reference="eq:Lvsf"}), where now \]\gamma=(K_ld)^{-1}\. \label{eq:gamma}\[ For the considered temperature, we find \(b_0\approx6.3\,\mu{\rm m}\) and \(K_l\approx10\,\mu{\rm m}\), so that formula ([\[eq:Lvsf\]](#eq:Lvsf){reference-type="ref" reference="eq:Lvsf"}) with \(\gamma\) defined by Eq. ([\[eq:gamma\]](#eq:gamma){reference-type="ref" reference="eq:gamma"}) provides a more accurate estimate yielding values of \(L\) which, for the same value of Andreev reflection \(f_R\), are about twice smaller than those predicted by the original model of Bradley *et al.* (however, note our comment in Ref. ).
# Spectral properties of Andreev reflection vs those of quantum turbulence {#sec:fluctuations}
In the simulation of quasiclassical turbulence using ring injection (see Fig. [\[fig:transient\]](#fig:transient){reference-type="ref" reference="fig:transient"} in Sec. [\[sec:vortgeneration\]](#sec:vortgeneration){reference-type="ref" reference="sec:vortgeneration"}), once the tangle has reached the statistically steady state, the vortex line density and the Andreev reflection coefficient fluctuate around their equilibrium, time-averaged values, \(\langle L\rangle=9.7\times10^{7}\,{\rm m}^{-2}\) and \(\langle f_R\rangle=0.34\), respectively. These fluctuations contain important information about the quantum turbulence. In order to compare the spectral characteristics of fluctuations of the Andreev reflection and the vortex line density, \]\delta f_R(t)=f_R(t)-\langle f_R\rangle\,, \quad \delta L(t)=L(t)-\langle L\rangle\,, \label{eq:delta-RL}\[ respectively, we monitor a steady state of simulated tangle for a period of approximately 380 s or 7500 snapshots. Taking the Fourier transform \(\widehat{\delta f_R}(f)\) of the time signal \(\delta f_R(t)\), where \(f\) is frequency, we compute the power spectral density \(\vert \widehat{\delta f_R}(f) \vert^2\) (PSD) of the Andreev reflection fluctuations. Similarly we compute the PSD \(\vert \widehat{\delta L}(f) \vert^2\) of the vortex line density fluctuations and plot both of them on Fig. [\[fig:anddensfluc\]](#fig:anddensfluc){reference-type="ref" reference="fig:anddensfluc"}. Figure [\[fig:anddensfluc\]](#fig:anddensfluc){reference-type="ref" reference="fig:anddensfluc"} also shows the PSDs of Andreev reflection fluctuations and of vortex line density fluctuations for mechanically driven (synthetic quasiclassical) and counterflow (ultraquantum) tangles. For each of the simulations, the PSDs of Andreev reflection and vortex line density fluctuations look very similar. A difference between the PSDs for our \(^3\)He tangle simulated using vortex ring injection is somewhat larger and may be caused by the presence of large scale flows that affect the spectral properties of the fluctuations, but not the time-averaged values of the Andreev reflection and vortex line densities. Our observations on the similarities of the PSD of Andreev reflection and vortex line density fluctuations are supported by the normalized cross-correlations between the fluctuations of the vortex line density, \(\delta L(t)\), and the fluctuations of the Andreev reflection, \(\delta f_R(t)\), which is computed as \]F_{LR}(\tau)=\frac{\langle \delta L(t)\delta f_R(t+\tau) \rangle} {\sqrt{\langle\delta L^2(0) \rangle}\sqrt{\langle\delta f_R^2(0) \rangle}}\,, \label{eq:correl}\(\) where the angle brackets indicate averaging over time, \(t\), in the saturated regime, and \(\tau\) is the time-lag. Figure [\[fig:cross-correlation\]](#fig:cross-correlation){reference-type="ref" reference="fig:cross-correlation"} shows the cross-correlation between the vortex line density and the Andreev reflection for the simulated tangles. The cross-correlation values \(F_{LR}(0)\) for ultraquantum and synthetically generated tangles approach unity, while for \(^3\)He tangle is slightly lower, \(F_{LR}(0)\approx0.9\). The high values of the cross-correlation clearly demonstrate a strong link between the fluctuations of Andreev reflection and vortex line density. We will now analyze the frequency dependence of the PSDs of the vortex line density fluctuations shown in Fig. [\[fig:anddensfluc\]](#fig:anddensfluc){reference-type="ref" reference="fig:anddensfluc"}. At high frequencies, the vortex line density fluctuation spectrum for all of the simulated tangles exhibits \(f^{-5/3}\) scaling and is dominated by the contribution of unpolarized, random vortex lines . In the intermediate frequency range, only the quasiclassical \(^3\)He fluctuation spectrum shows a clear \(f^{-3}\) scaling, which is governed by the large scale flows indicating polarized (that is, cooperatively correlated) vortex lines in agreement with the recent numerical simulations . If we reasonably assume that a crossover between \(f^{-3}\) to the \(f^{-5/3}\) behavior should occur at around the frequency corresponding to the intervortex distance, \(\ell\), then using the value of 102\(\,\mu\)m for \(\ell\) calculated above for the equilibrium tangle, the crossover should occur at frequency \(f_{\ell} \approx v/\ell=\kappa/(2 \pi \ell)\approx1\) Hz. As seen in Fig. [\[fig:anddensfluc\]](#fig:anddensfluc){reference-type="ref" reference="fig:anddensfluc"}, this is in a very fair agreement with the frequency \(\approx2\) Hz of the crossover between the two regimes. Figure [\[fig:flucsimtanglerings\]](#fig:flucsimtanglerings){reference-type="ref" reference="fig:flucsimtanglerings"} shows and compares the PSDs of fluctuations of Andreev reflection for the simulation (top, blue) and for the experiments (middle and bottom, gray). The experimental data are shown for a fully-developed tangle (dark gray) and ballistic vortex rings (light gray) corresponding to the grid velocities of 6.3 mm s\(^{-1}\) and 1.9 mm s\(^{-1}\) respectively. (The prominent peaks in the numerical data are artifacts of the discrete vortex-ring injection process.) The power spectrum of \(\delta f_R(t)\) of the simulation and of the experimental data for the developed tangle (reported here and in Refs. ) are in overall agreement showing reduction of the signal with increasing frequency. A detailed comparison of experimental observations at high grid velocity and simulations can be based on two alternative approaches to the interpretation of numerical and experimental data. Both of these approaches are viable, mostly due to a rather noisy character of the fluctuating Andreev reflection's spectra for simulated tangles, but they lead to somewhat different conclusions. In the first approach we could identify the same \(f^{-5/3}\) scaling behavior at intermediate frequencies. This scaling, observed experimentally in high temperature \(^4\)He and reproduced by numerical simulations, has been interpreted as that corresponding to the fluctuations of passive (material) lines in Kolmogorov turbulence. At high frequencies the experimental data develop a much steeper scaling (\(\approx\,f^{-3}\)), not seen in the numerical spectrum, probably owing to the finite numerical resolution. However, this frequency dependence *is* observed in the experiment in the case where only microscopic vortex rings are propagating through the active region and there are no large-scale flows or structures. Thus, we can argue that the \(f^{-3}\) scaling for the vortex tangles corresponds to Andreev reflection from superflows on length scales smaller than the intervortex distance. In the second, alternative approach we assume that the PSD of Andreev reflection fluctuations mimics the functional dependence of the PSD of the vortex line density in accordance with Fig. [\[fig:flucsimtanglerings\]](#fig:flucsimtanglerings){reference-type="ref" reference="fig:flucsimtanglerings"}. In this case the Andreev-reflection fluctuation spectra exhibit a \(f^{-5/3}\) dependence at high frequency and a steeper (\(\approx\,f^{-3}\)) dependence at intermediate frequencies, which is almost opposite to the scaling observed in the experiment. Figure [\[fig:polarised\]](#fig:polarised){reference-type="ref" reference="fig:polarised"} shows fluctuation spectra of the vortex line density calculated for the total line density (black upper solid line) and the polarized fraction of the quasiclassical tangle (red lower solid line). The observed frequency dependencies are similar, and suggest that in our experiments the vortices probed by quasiparticles are polarized and hence the tangle generated by the vibrating grid is much more polarized than expected. It is feasible that vortex rings moving predominantly in the same direction polarize the resulting tangle, and that quasiparticles accompanying turbulence production further assist the polarization of the tangle in the direction of grid motion.
# Discussion and conclusions {#sec:interpretations}
In this study we combined the three-dimensional numerical simulations of dynamics and evolution of turbulent vortex tangles, numerical analysis of Andreev scattering of thermal quasiparticle excitations by the simulated tangles, and experimental measurements of Andreev reflection by quantum turbulence in \(^3\)He-B at ultralow temperatures. Our aim was to show that a combined study is a powerful tool that might reveal important properties of pure quantum turbulence. We have shown that, with sufficient resolution of the Andreev retroreflected signal, the Andreev scattering technique can generate a map of local reflections which reveals a structure of the vortex tangle. Furthermore, in the numerical experiment it is possible to separate the quasiparticle and quasihole excitations. The individual maps of Andreev reflection for each of these species of excitations also show the presence and the structure of large scale flows arising in the quasiclassical (Kolmogorov) regime of quantum turbulence. However, in the real physical experiment the thermal beam consists of excitations of both sorts whose separation is impossible. We found that, while a presence of large scale flows may still be detected by the enhancement of local Andreev reflections, an identification of the detailed large scale flow structure would hardly be possible. For the quasiclassical vortex tangle, we calculated the total Andreev reflection as a function of the vortex line density. We found that the results of our numerical simulation reproduce rather well the experimental measurements of the total Andreev reflection, not only qualitatively but also quantitatively, although this good quantitative agreement should be taken with some caution for the reason explained in Sec. [\[sec:exptl\]](#sec:exptl){reference-type="ref" reference="sec:exptl"}. We also found that the total reflection is strongly affected by the screening effects. Here we use the term "screening" to identify mechanisms which, for a given line density, change the overall reflectivity of the tangle. In particular, screening mechanisms are responsible for the slowing of growth of the total reflectivity with the vortex line density when the latter reaches, in our simulation, the value about \(2\times10^7\,{\rm m}^{-2}\). We identified two main mechanisms of screening. The first, so-called "partial screening" may significantly reduce the total reflectivity of the dense vortex configuration due to a modification of the \(1/r\) superflow in the vicinity of the vortex line. The second is the 'fractional' (or 'geometric') screening and is of importance for dense or/and large tangles: Vortices located closer to the source of excitations screen those at the rear. We investigated screening mechanisms in some detail and conclude that they often should be taken into consideration in interpretation of experimental observations and numerical simulations. While this paper is mostly concerned with the Andreev reflection from quasiclassical vortex tangles, there also exists another regime of quantum turbulence of a very different spectral nature in which the energy is contained in the intermediate range of scales and the energy spectrum nowhere conforms to the Kolmogorov \(-5/3\) scaling. Unlike quasiclassical quantum turbulence, this form of turbulence, named 'ultraquantum' or 'Vinen' turbulence , is unstructured: the only length scale in the ultraquantum tangle is the mean intervortex separation. At very low temperatures the ultraquantum regime of quantum turbulence has been identified experimentally in \(^4\)He and \(^3\)He-B . Considering that the Andreev reflection can reveal the presence of the large scale flows existing in quasiclassical turbulence, it could be expected that the ultraquantum tangle can be distinguished from the quasiclassical one by their total Andreev reflectivities. However, our modeling study in Sec. [\[sec:ultraquantum\]](#sec:ultraquantum){reference-type="ref" reference="sec:ultraquantum"} of Andreev reflection from the structured (Kolmogorov) and unstructured (ultraquantum) vortex tangles for the same vortex line density has demonstrated that this is not the case, and that the total reflection is determined by the vortex line density alone. This, in particular, means that the contribution of the large scale flow to the total Andreev reflection is small. Finally, we investigated the spectral characteristics of Andreev reflection from various tangles and compared them with the spectral characteristics of the quantum turbulence. We found that the fluctuations of the vortex line density and of the Andreev reflection are very strongly correlated and also have similar fluctuation power spectra. The Andreev-reflection fluctuation spectra and of the vortex line density fluctuations look practically identical in the case of ultraquantum turbulence but may differ for the quasiclassical tangle, that is when a large scale flow is present. Therefore, some caution should be taken in interpreting the \(f^{-5/3}\) scaling of the Andreev retroreflected signal's frequency spectrum, observed in the experiment of Bradley *et al.* , as a direct representation of the spectral properties of the vortex-line density fluctuations in turbulent \(^3\)He-B. Our results show that this scaling could be attributed to the fluctuations of passive (material) lines in Kolmogorov turbulence, or to the fluctuations of the vortex line density of a strongly polarized tangle. The latter possibility suggests that a turbulent tangle produced by the vibrating grid at high velocities is more polarized than has usually been assumed earlier. It is very clear that the Andreev reflection technique has great potential for elucidating the properties of pure quantum turbulence, but requires further work. It is clear that a useful study could be made of a tangle created from rings emitted by two grids facing each other. Such a tangle should have a vortex line density similar to that of the fully-developed tangle studied in our work but with a much smaller polarization. Numerically, it would be interesting to investigate fluctuations of the Andreev reflection and vortex line density for the shower of ballistic vortex rings using a higher spatial resolution and whether the \(f^{-3}\) spectrum would reproduced experimentally. The influence of large scale flows on Andreev reflections should also be studied in superfluid \(^3\)He under rotation since this scenario uniquely offers a vortex line ensemble with inherent strong polarization. | {'timestamp': '2017-06-08T02:06:07', 'yymm': '1706', 'arxiv_id': '1706.01791', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01791'} |
# Introduction
## Mahler's conjecture for the volume product
A *convex body* in \(\R^n\) is a compact convex set in \(\R^n\) with nonempty interior. Denote by \(\K^n\) the set of all convex bodies in \(\R^n\). A convex body \(K \in \K^n\) is said to be *centrally symmetric* if it satisfies \(K=-K\). We denote by \(\K^n_0\) the set of all \(K \in \K^n\) which are centrally symmetric. Let \(K \in \K^n\) be a convex body. The interior of \(K\) is denoted by \(\interior K\). For \(z \in \interior K\), the *polar body* of \(K\) with respect to \(z\) is defined by \[K^z = \left\{ y \in \R^n; (y-z) \cdot (x-z) \leq 1 \text{ for any } x \in K \right\},\] where \(\cdot\) denotes the standard inner product on \(\R^n\). Denote by \(|K|\) the \(n\)-dimensional volume of \(K\) in \(\R^n\). Then the *volume product* of \(K\) is defined by \[\label{eq:a} \mathcal{P}(K) := \min_{z \in \interior K}|K| \, |K^z|.\] Note that this quantity is an affine invariant, i.e., \(\mathcal{P}(\linearop K)=\mathcal{P}(K)\) for any invertible affine map \(\linearop :\R^n \to \R^n\). It is well-known that for each \(K \in \K^n\) the minimum of [\[eq:a\]](#eq:a){reference-type="eqref" reference="eq:a"} is attained at the unique point \(z\) on \(K\), which is called *Santaló point* of \(K\) (see ). For a centrally symmetric convex body \(K \in \K^n_0\), the Santaló point of \(K\) is nothing but the origin \(O\). In the following, the polar of \(K\) with respect to \(O\) is denoted by \(K^\circ\). The upper bound for \(\mathcal{P}(K)\) is well-known as Blaschke-Santaló inequality (see ) and the bound is attained only for ellipsoids (see ). In contrast, the sharp lower bound estimate is fairly difficult and it remains as a longstanding open problem since 1939 as follows.
This is trivial for \(n=1\) and was proved by Mahler himself for \(n=2\). There are many alternative proofs and the characterization of the equality case for \(n=2\) (see, e.g.,, \*Section 10.7, and references therein). However, the case where \(n \geq 3\) is still widely open. We solve affirmatively symmetric Mahler's conjecture in the three dimensional case.
Further, we exhibit known partial results about the conjecture. An asymptotic lower bound of the volume product was first given by Bourgain and Milman and the best known constant is due to Kuperberg. Meanwhile, the conjecture itself has been proved for very restricted cases; for zonoids by Reisner [@R; @R2] (see also ) and for unconditional bodies by Saint-Raymond. Note that the equality of [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"} is attained by an \(n\)-cube, for instance. In, the estimate [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"} was confirmed for \(K \in \K_0^n\) sufficiently close to the unit \(n\)-cube in the Banach-Mazur distance. For other partial results, see e.g.,, , ,. There is another question when \(K \in \K^n\) is not necessarily assumed to be centrally symmetric.
This was proved by Mahler for \(n=2\) (see ). The alternative proof of Mahler's result by Meyer is notable. Indeed, we were able to extract an important idea for attacking to the three dimensional symmetric case from. We first improve and simplify the idea and adapt it to the symmetric case (see Section [2](#sec:2){reference-type="ref" reference="sec:2"}). This method is applicable to not only the two dimensional case but also the three dimensional case. Note that non-symmetric Mahler's conjecture remains open for \(n \geq 3\), see e.g.,, , ,.
## An application to Viterbo's conjecture
Recently, a surprising connection between symmetric Mahler's conjecture and a conjecture in the field of symplectic geometry was discovered. In 2000, Viterbo posed in an isoperimetric-type conjecture for symplectic capacities of convex bodies in \(\R^{2n}\) with the standard symplectic structure \(\omega_0\). A symplectic capacity \(c\) is a symplectic invariant which assigns a non-negative real number to each of symplectic manifolds of dimension \(2n\). The Hofer-Zehnder capacity \(c_{\rm HZ}\) is one of the important symplectic capacities, which is related to Hamiltonian dynamics on symplectic manifolds. For details, see or a foundational book.
This conjecture is unsolved even in the case of \(n=2\). Note that \(c(B^{2n})=\pi\). In, Artstein-Avidan, Karasev, and Ostrover calculated the Hofer-Zehnder capacity of a convex domain \(K \times K^\circ \subset \R^{2n}\) \((K \in \K_0^n)\) as follows.
Using it, they gave the following remarkable observation that *Viterbo's conjecture implies symmetric Mahler's conjecture* (see \*Section 1). \[\frac{4^n}{\pi^n} = \frac{c_{\rm HZ}(K \times K^\circ)^n}{\pi^n} \leq \frac{\mathrm{vol}(K \times K^\circ)}{\pi^n/n!} = \frac{\abs{K} \abs{K^\circ}}{\pi^n/n!}.\] Conversely, if we assume that symmetric Mahler's conjecture holds, then Theorem [\[thm:a\]](#thm:a){reference-type="ref" reference="thm:a"} implies the inequality [\[eq:c\]](#eq:c){reference-type="eqref" reference="eq:c"} for the case that \(c=c_{\rm HZ}\) and \(\Sigma=K \times K^\circ\). Therefore, Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} immediately yields
Of course, due to the result of , the inequality [\[eq:c\]](#eq:c){reference-type="eqref" reference="eq:c"} also holds for \(\Sigma=K \times K^\circ \subset (\R^4,\omega_0)\), where \(K \in \K_0^2\), with respect to \(c_{\rm HZ}\).
## Organization of the paper and our method
Our approach to the conjecture started with a trial to adapt the proof of the two dimensional non-symmetric case () to the symmetric case. We then found a key idea to estimate \(\mathcal{P}(K)\) effectively. The case where \(n=2\) contains the essence of the idea in a primitive way. Moreover that is useful to understand the strategy of our proof of the three dimensional case. Let \(K \in \K_0^2\). First, by using coordinate axises, we divide \(K\) into four parts \(K_1, K_2, K_3(=-K_1), K_4(=-K_2)\). Here we may assume that \(|K_1|=|K_2|\). Next, we consider the corresponding decomposition of \(K^\circ\). Roughly speaking, the division \(K^\circ=K^\circ_1 \cup K^\circ_2 \cup-K^\circ_1 \cup-K^\circ_2\) is obtained by using the following two lines. One is the line through the origin \(O\) and a point \(B^\circ\) on \(\partial K^\circ\) with the maximal \(x\)-coordinate. Another is the line through \(O\) and a point \(C^\circ\) on \(\partial K^\circ\) with the maximal \(y\)-coordinate. Then, for any point \(P \in K^\circ\), the areas \(|K^\circ_i|\) \((i=1,2)\) are estimated from below by the signed area of the quadrilaterals \(OB^\circ PC^\circ\) and \(OC^\circ P(-B^\circ)\), respectively. These estimates give two test points \(S_1, S_2\) on \(K\). The same observation on \(K\) yields further two test points \(R_1, R_2\) on \(K^\circ\). By paring \(R_i\) and \(S_i\) \((i=1,2)\), we easily get \(\mathcal{P}(K) \geq 8\). Including the equality case we exhibit the details in Section [2](#sec:2){reference-type="ref" reference="sec:2"}. From Section [3](#sec:3){reference-type="ref" reference="sec:3"}, we begin the proof of the three dimensional case. We first divide \(K\) into eight parts \(K_i\) \((i=1,\ldots,8)\) as in the two dimensional case. However, compare to the two dimensional case, to make the corresponding decomposition of \(K^\circ\) is not so simple. The first ingredient of our proof of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} is to concentrate only on the class of convex bodies \(K \in \K_0^3\) which are *strongly convex and \(C^\infty\) boundary*. Schneider's approximation procedure (Proposition [\[prop:13\]](#prop:13){reference-type="ref" reference="prop:13"}) guarantees that the problem is reduced to considering only this class. For a convex body \(K\) in the class, we can define a diffeomorphism \(\partial K \to \partial K^\circ\), which yields the corresponding eight pieces decomposition \(K_i^\circ\) of \(K^\circ\) from that of \(K\). The precise setting is explained in Section [3.2](#sec:3.2){reference-type="ref" reference="sec:3.2"}. The second ingredient is that only a few test points give a sharp estimate of \(\mathcal{P}(K)\), *provided \(K\) has an additional symmetry*. When \(n=2\), the symmetry we need is \(|K_1|=|K_2|\), although this is not an essential assumption. When \(n=3\), we have to control not only the volume of \(K_i\) but also the areas of boundary faces of \(K_i\) in the three coordinate planes. In Section [3.5](#sec:3.5){reference-type="ref" reference="sec:3.5"}, we give a sharp estimate of \(\mathcal{P}(K)\) from below under the condition that \(K\) has "good" symmetries. The condition is [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}. Under this condition, the process of estimating \(\mathcal{P}(K)\) is regarded as a direct generalization of the two dimensional case. A (signed) volume comparison inequality (Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"}) which yields test points on \(K\) and \(K^\circ\) is in Section [3.3](#sec:3.3){reference-type="ref" reference="sec:3.3"}, of which proof is given in Section [3.4](#sec:3.4){reference-type="ref" reference="sec:3.4"}. How can we release \(K\) from the very strong condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}? Note that \(\mathcal{P}(K)\) is invariant under linear transformations of \(\R^3\). The third ingredient is to make the most of the freedom of the action of linear transformations. As a test case, in Section [4](#sec:4){reference-type="ref" reference="sec:4"}, we prove Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} for \(K \in \K^3_0\) equipped with an extra symmetry with respect to a hyperplane through \(O\) (Proposition [\[prop:15\]](#prop:15){reference-type="ref" reference="prop:15"}). In this case, by a linear transformation \(\mathcal A\), we can deform \(K\) into \(\mathcal{A}K\) which satisfies the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} by means of the intermediate value theorem. However, for a general \(K\) to find a linear transformation \(\mathcal A\) such that \(\mathcal{A}K\) satisfies the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} is highly nontrivial. Sections [5](#sec:5){reference-type="ref" reference="sec:5"} and [6](#sec:6){reference-type="ref" reference="sec:6"}, which are the technical part of the present paper, are devoted to find an appropriate linear transformation \(\mathcal A\) for each \(K \in \K^3_0\) which is strongly convex with smooth boundary \(\partial K\). We proceed as follows. First, we define \(K(\theta, \phi, \psi) \in \K^3_0\) as the image of the initial \(K\) under the action of \(SO(3)\). (Here, \(\theta, \phi, \psi\) mean rotation angles with \(x, y, z\)-axises, respectively.) Next, in Section [5.2](#sec:5.2){reference-type="ref" reference="sec:5.2"} we define a linear transform \(\mathcal A\) such that \(\mathcal{A}K (\theta, \phi, \psi)\) satisfies the condition [\[eq:22\]](#eq:22){reference-type="eqref" reference="eq:22"} (Proposition [\[prop:3\]](#prop:3){reference-type="ref" reference="prop:3"}), which is a partial condition of [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}. Finally, to get the full condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} for \(\mathcal{A}K (\theta, \phi, \psi)\), we introduce in Section [5.3](#sec:5.3){reference-type="ref" reference="sec:5.3"} three smooth functions \(F(K), G(K), H(K)\) on a contractible region \(D \subset \mathbb R^3\) with the coordinates \(\theta, \phi, \psi\). These functions are defined by the volume of pieces of the eight part decomposition of \(\mathcal{A}K(\theta, \phi, \psi)\) and the resulting two dimensional quantities. Then the problem to find \(\mathcal{A}K\) which satisfies the desired condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} is reduced to the existence of a zero \((\theta, \phi, \psi)\) of the map \((F,G,H): D \to \mathbb R^3\). Here, if \((F,G,H)\) has no zero on \(\partial D\), then we can define a map \[\mathscr{F}=\frac{(F,G,H)}{\sqrt{F^2+G^2+H^2}}:\partial D \to S^2.\] From Section [5.4](#sec:5.4){reference-type="ref" reference="sec:5.4"} to [5.8](#sec:5.8){reference-type="ref" reference="sec:5.8"}, we discuss properties of the map \((F,G,H)\) in order to calculate the degree of the map \(\mathscr{F}\). The centrally symmetric convex body \(\mathcal{A}K(\theta, \phi, \psi)\) equips with additional symmetries under rotations by specific angles around \(x,y,z\)-axises. These induces certain symmetries of the functions \(F,G\), and \(H\) (Lemmas [\[lem:5\]](#lem:5){reference-type="ref" reference="lem:5"}, [\[lem:6\]](#lem:6){reference-type="ref" reference="lem:6"}, [\[lem:8\]](#lem:8){reference-type="ref" reference="lem:8"}, and [\[lem:10\]](#lem:10){reference-type="ref" reference="lem:10"}). The results are summarized in Proposition [\[prop:7\]](#prop:7){reference-type="ref" reference="prop:7"} and essential for the degree calculation in the next section. In Section 6, we actually prove that \(\deg \mathscr{F} \neq 0\) after a suitable perturbation of this map \(\mathscr{F}\). This immediately implies the existence of a zero of the map \((F,G,H)\). Note that the smoothness of \(\partial K\) also has enough merit to the calculation of \(\deg \mathscr{F}\). The construction of the above perturbation of \(\mathscr{F}\) is technical, so it is discussed in Appendix [\[sec:A\]](#sec:A){reference-type="ref" reference="sec:A"}. Anyway the perturbed \(\mathscr{F}\) equips with \((\pm1,0,0)\) as regular values. This fact is crucial for the calculation of \(\deg \mathscr{F}\). Indeed, it enables us to reduce the calculation of \(\deg \mathscr{F}\) to the calculation of the winding number of a map \(\mathscr{G}: S^1 \to S^1\) (Proposition [\[prop:5\]](#prop:5){reference-type="ref" reference="prop:5"}), which is fairly accessible. We see that the winding number of \(\mathscr{G}\) is odd. This reduction process is explained in Section [6.2](#sec:6.2){reference-type="ref" reference="sec:6.2"} and the calculation of the winding number of the map \(\mathscr{G}\) is carried out in Section [6.3](#sec:6.3){reference-type="ref" reference="sec:6.3"}. Consequently, the existence of a zero of the map \((F,G,H)\) ensures that \(\mathcal{P}(K) \geq 32/3\) for any convex body \(K \in \K_0^3\) which is strongly convex with \(C^\infty\) boundary. Finally, Schneider's approximation theorem implies that the inequality holds for all \(K \in \K_0^3\). In Section [7](#sec:7){reference-type="ref" reference="sec:7"}, we determine the equality condition (Propositions [\[prop:10\]](#prop:10){reference-type="ref" reference="prop:10"} and [\[prop:11\]](#prop:11){reference-type="ref" reference="prop:11"}). First, for a general \(K \in \K^3_0\) with \(\mathcal{P}(K)=32/3\), we find a linear transformation \(\mathcal{A}\) such that \(\mathcal{A}K\) satisfies the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}. Indeed, there exists a sequence \(\{K_n\}_{n \in \N}\) such that \(\lim_{n \rightarrow \infty} K_n=K\), each \(K_n \in \K^3_0\) is strongly convex with smooth boundary. By the results in Sections [5](#sec:5){reference-type="ref" reference="sec:5"} and [6](#sec:6){reference-type="ref" reference="sec:6"}, there exists \(\{\mathcal{A}_n\}_{n \in \N}\) such that each \(\mathcal{A}_n K_n\) satisfies the condition \(\eqref{eq:3}\). As the limit of a subsequence, we get the required \(\mathcal{A}\). Moreover, from \(\mathcal{P}(K)=32/3\), we can show that \(P_i(K^\circ)=(Q_i(K))^\circ\) and \(\abs{Q_i(K)} \abs{P_i(K^\circ)}=8\) where \(Q_i\) is the cross section with a coordinate plane and \(P_i\) is the projection to the coordinate plane \((i=1,2,3)\). Thus, the result of the two dimensioal case (Proposition [\[prop:9\]](#prop:9){reference-type="ref" reference="prop:9"}) implies that \(Q_i(K)\) and \(P_i(K^\circ)\) are parallelograms. Next, we extend the sharp estimate in Section [3.5](#sec:3.5){reference-type="ref" reference="sec:3.5"}. Then we see \(K\) is polyhedron with at most 20 vertices. Finally, by case analysis and calculating the dual faces of the vertices, we prove that either \(K\) or \(K^\circ\) is a parallelepiped.
# The two dimensional case {#sec:2}
In this section, we give a simple proof of Mahler's theorem for the two dimensional case. The essence of the idea originates from. The goal of this section is the following
To prove it, for \(K \in \K_0^2\), we divide \(K\) into the following four pieces: \[\begin{aligned}
K_1 &:= K \cap \left\{ (x,y) \in \R^2; x \geq 0, y \geq 0 \right\}, \\ K_2 &:= K \cap \left\{ (x,y) \in \R^2; x \leq 0, y \geq 0 \right\}, \\ K_3 &:=-K_1= K \cap \left\{ (x,y) \in \R^2; x \leq 0, y \leq 0 \right\}, \\ K_4 &:=-K_2= K \cap \left\{ (x,y) \in \R^2; x \geq 0, y \leq 0 \right\}. \end{aligned}\] We denote by \(\polygon \left\{P_1, \dots, P_m\right\}\) the star-shaped polygon in \(\R^2\) consists of successive vertices \(P_1, \dots, P_m \, (m \geq 3)\) in the counterclockwise order and the edges \(P_1 P_2, \dots, P_{m-1} P_m, P_m P_1\).
# A sharp estimate for the three dimensional volume product {#sec:3}
## Preliminaries {#sec:3.1}
To begin with, we summarize some notations and terminologies about convex bodies in arbitrary dimension. Let \(K \in \K^n\) be an \(n\)-dimensional convex body with \(O \in \interior K\). The *Minkowski gauge* \(\mu_K\) of \(K\) is defined by \[\mu_K(x) := \min\{ \lambda \geq 0; x \in \lambda K \}\] for \(x \in \R^n\), and the *radial function* \(\rho_K\) of \(K\) is defined by \[\rho_K(x) := \max\{ \lambda \geq 0; \lambda x \in K \} = \frac{1}{\mu_K(x)}\] for \(x \in \R^n \setminus \{ O \}\). Note that \(\rho_K(x)x \in \partial K\) for any \(x \in \R^n \setminus \{ O \}\). For a point \(x \in \R^n\), denote by \(-x\) the opposite point of \(x\) with respect to the origin \(O\). And we abbreviately use the same symbol \(x\) to represent the position vector of \(x\). For a curve \(c(t) \ (0 \leq t \leq 1)\) in \(\R^n\), denote by \(\widetilde{c}\) the curve which has the same image as \(c\) with the reverse orientation, that is, \[\widetilde{c}(t) = c(1-t) \quad (0 \leq t \leq 1).\] Let \(c,d\) be smooth curves in \(\R^n\). Suppose that the terminal point of \(c\) coincides with the initial point of \(d\). Then denote by \(c \cup d\) the piecewise smooth curve in \(\R^n\) which starts at \(c(0)\) and ends at \(d(1)\). Now we examine the correspondence of the boundary points between a convex body \(K\) and its polar body \(K^\circ\), when \(K\) is a strongly convex body defined below. Assume that the boundary \(\partial K\) is a \(C^\infty\)-hypersurface in \(\R^n\). Consider a smooth function \[h_K(x) = \frac{1}{2} \mu_K^2(x)\] on \(\R^n\). The condition that \(x \in \partial K\) is equivalent to \(h_K(x) = 1/2\), and hence \(\partial K\) is the level set \(h_K^{-1}(1/2)\) and the gradient \(\nabla h_K(x)\) is an outward normal vector at \(x \in \partial K\). We say that \(K\) is *strongly convex* if \(h_K\) is a strongly convex function, that is, the Hessian matrix \(D^2 h_K(x)\) is positive definite for each \(x \in S^{n-1}=\left\{ x \in \R^n; |x|=1 \right\}\). Assume that \(K\) is strongly convex. By the definition of polar, for every \(y \in \partial K^\circ\) there exists the unique point \(x \in \partial K\) such that \(y \cdot x = 1\). This correspondence gives rise to the map \(\hat{\Lambda}: \partial K^\circ \rightarrow \partial K\) defined by \(\hat{\Lambda}(y):=x\). Since the gradient \(\nabla h_K(x)\) at the point \(x\) is the same direction as \(y\), we have \[\nabla h_K(x) = a(y) y\] with a positive function \(a\) on \(\partial K^\circ\). Since \(\nabla h_K\) is \(1\)-homogeneous, Euler's formula implies \[1= y \cdot x = \frac{\nabla h_K(x)}{a(y)} \cdot x = \frac{2 h_K(x)}{a(y)}.\] This means that the condition \(x \in \partial K\) is equivalent to \(a(y)=1\), so that \[\nabla h_K(x) = y \ \text{ for } \ y \in \partial K^\circ.\] From \(\hat{\Lambda}(y)=x\), we find that a smooth map \(\nabla h_K \rvert_{\partial K}: \partial K \to \partial K^\circ\) is the inverse of \(\hat{\Lambda}\). We define \(\Lambda:= \nabla h_K \rvert_{\partial K}\).
Note that by [\[eq:56\]](#eq:56){reference-type="eqref" reference="eq:56"}, \(K^\circ\) is smooth strongly convex if \(K\) is smooth strongly convex.
## Convex bodies in \(\R^3\) {#sec:3.2}
From now on, we focus on centrally symmetric convex bodies in \(\R^3\). Denote by \(\hat{\K}\) the set of all centrally symmetric convex bodies \(K \in \K^3_0\) which are strongly convex with smooth boundary \(\partial K\). Throughout this paper, for a convex body \(K \in\) \(\K^3\) we fix the orientation on \(K\) induced from the natural orientation of \(\R^3\). From it the orientation of the boundary \(\partial K\), that of any domain \(S\) on \(\partial K\) and that of \(\partial S\) (that is, a closed curve on \(\partial K\)) are induced, respectively. For a point \(P=(p_1,p_2,p_3)\) in \(\R^3\) we write \(-P=(-p_1,-p_2,-p_3)\). If \(K \in \hat{\K}\) and \(P \in K\), then \(-P \in K\). For \(K \in \hat{\K}\) and any distinct two points \(P, Q \in \partial K\) with \(P \not=-Q\), let us introduce an oriented curve from \(P\) to \(Q\) on the boundary \(\partial K\) defined by \[\mathcal{C}_K(P,Q) := \left\{ \rho_K((1-t) P + t Q) ((1-t)P + t Q); \ 0 \leq t \leq 1 \right\}.\] We call it an *oriented segment* on \(\partial K\). The polar body of \(K \in \hat{\K}\) is given by \[K^\circ = \left\{ (u,v,w) \in \R^3; (u,v,w) \cdot (x,y,z) \leq 1 \text{ for all } (x,y,z) \in K \right\},\] which is also an element of \(\hat{\K}\) from the argument in Section [3.1](#sec:3.1){reference-type="ref" reference="sec:3.1"}. Note that \(O \in \interior K^\circ\) and \((K^\circ)^\circ=K\). Let \(K \in \hat{\K}\). First, we decompose the convex body \(K\) into eight pieces. Using the radial function \(\rho_{K}\) we take three points \[A:=\rho_{K}((1,0,0)) (1,0,0), B:=\rho_{K}((0,1,0)) (0,1,0), C:=\rho_{K}((0,0,1)) (0,0,1)\] on the boundary \(\partial K\). These points lie in the \(x\)-axis, \(y\)-axis, and \(z\)-axis, respectively, so we can write \(A=(a_1,0,0), B=(0,b_2,0), C=(0,0,c_3)\) with \(a_1, b_2, c_3>0\). Since \(K\) is centrally symmetric, we have \(-A,-B,-C \in \partial K\). We consider six oriented segments on \(\partial K\) as follows: \[\begin{aligned}
d&:=\mathcal{C}_{K}(B, C), & e&:=\mathcal{C}_{K}(C,-B), \\ f&:=\mathcal{C}_{K}(C, A), & g&:=\mathcal{C}_{K}(A,-C), \\ h&:=\mathcal{C}_{K}(A, B), & i&:=\mathcal{C}_{K}(B,-A). \end{aligned}\] Then another six segments \(-d,-e,-f,-g,-h\), and \(-i\) on \(\partial K\) are automatically defined. Note that \(\pm d,\pm e\) are in the \(yz\)-plane, \(\pm f,\pm g\) are in the \(zx\)-plane, and \(\pm h,\pm i\) are in the \(xy\)-plane. We denote by \(Q_i(K)\ (i=1,2,3)\) the sections of \(K\) by the \(yz\)-plane, \(zx\)-plane, and \(xy\)-plane, respectively. They are strongly convex sets in each plane and closed curves \(d \cup e \cup-d \cup-e\), \(f \cup g \cup-f \cup-g\), and \(h \cup i \cup-h \cup-i\) are the boundaries of \(Q_i(K)\ (i=1,2,3)\), respectively. Hence these closed curves mutually do not intersect except for \(\pm A, \pm B, \pm C\). For an oriented simple closed curve \(c\) on \(\partial K\) denote by \(\mathcal{S}_K(c) \subset \partial K\) the piece of surface enclosed by \(c\) with the orientation compatible with that of \(c\). In this paper, we always equip \(\mathcal{S}_K(c)\) with this orientation. We denote by \(O*\mathcal{S}_K(c)\) the cone over \(\mathcal{S}_K(c)\): \[O*\mathcal{S}_K(c) = \{ \lambda u; u \in \mathcal{S}_K(c) \text{ and } 0 \leq \lambda \leq 1 \}.\] Using this notation, we divide \(K\) into the following eight pieces: \[\begin{aligned}
K_1&:=O*\mathcal{S}_K(d \cup f \cup h), & K_2&:=O*\mathcal{S}_K(\widetilde{d}\, \cup i \cup-g), \\ K_3&:=O*\mathcal{S}_K(\widetilde{e}\, \cup \widetilde{-g}\, \cup-h), & K_4&:=O*\mathcal{S}_K(e \cup-i \cup \widetilde{f}), \\ K_5&:=O*\mathcal{S}_K(-e \cup \, \widetilde{h}\, \cup g), & K_6&:=O*\mathcal{S}_K(\widetilde{-e}\, \cup-f \cup \widetilde{i}), \\ K_7&:=O*\mathcal{S}_K(\widetilde{-d}\, \cup \widetilde{-h}\, \cup \widetilde{-f}), & K_8&:=O*\mathcal{S}_K(-d \cup \, \widetilde{g}\, \cup \widetilde{-i}). \end{aligned}\] The volume of each \(K_i\ (i=1,2,\ldots,8)\) is denoted by \(|K_i|\). Next, we decompose the polar body \(K^\circ\) into eight pieces associated to the above decomposition of \(K\). Using the map \(\Lambda: \partial K \to \partial K^\circ\) defined in Section [3.1](#sec:3.1){reference-type="ref" reference="sec:3.1"}, we put three points and six curves on \(\partial K^\circ\) as follows: \[\begin{aligned}
&A^\circ := \Lambda(A), \quad B^\circ := \Lambda(B), \quad C^\circ := \Lambda(C); \\ &d^\circ:= \Lambda(d), \quad e^\circ:= \Lambda(e), \quad f^\circ:= \Lambda(f), \quad g^\circ:= \Lambda(g), \quad h^\circ:= \Lambda(h), \quad i^\circ:= \Lambda(i). \end{aligned}\] From the definition of the polar body and the strong convexity of \(K^\circ\), the point \(A^\circ\) (resp. \(B^\circ, C^\circ\)) is the unique point with the maximal \(u\)-coordinate (resp. \(v\)-coordinate, \(w\)-coordinate) among all the points in \(K^\circ\). Since \(K^\circ\) is centrally symmetric, we have \(-A^\circ,-B^\circ,-C^\circ \in \partial K^\circ\) and \(-d^\circ,-e^\circ,-f^\circ,-g^\circ,-h^\circ,-i^\circ \subset \partial K^\circ\). Since \(\Lambda: \partial K \to \partial K^\circ\) is a \(C^{\infty}\)-diffeomorphism, smooth closed curves \(d^\circ \cup e^\circ \cup-d^\circ \cup-e^\circ\), \(f^\circ \cup g^\circ \cup-f^\circ \cup-g^\circ\), and \(h^\circ \cup i^\circ \cup-h^\circ \cup-i^\circ\) on \(\partial K^\circ\) are mutually do not intersect except for \(\pm A^\circ, \pm B^\circ, \pm C^\circ\). Thus we can define a decomposition of \(K^\circ\) into the following eight pieces \[\begin{aligned}
K^\circ_1&:=O*\mathcal{S}_{K^\circ}(d^\circ \cup f^\circ \cup h^\circ), & K^\circ_2&:=O*\mathcal{S}_{K^\circ}(\widetilde{d^\circ} \cup i^\circ \cup-g^\circ), \\ K^\circ_3&:=O*\mathcal{S}_{K^\circ}(\widetilde{e^\circ} \cup \widetilde{-g^\circ} \cup-h^\circ), & K^\circ_4&:=O*\mathcal{S}_{K^\circ}(e^\circ \cup-i^\circ \cup \widetilde{f^\circ}), \\ K^\circ_5&:=O*\mathcal{S}_{K^\circ}(-e^\circ \cup \widetilde{h^\circ} \cup g^\circ), & K^\circ_6&:=O*\mathcal{S}_{K^\circ}(\widetilde{-e^\circ} \cup-f^\circ \cup \widetilde{i^\circ}), \\ K^\circ_7&:=O*\mathcal{S}_{K^\circ}(\widetilde{-d^\circ} \cup \widetilde{-h^\circ} \cup \widetilde{-f^\circ}), & K^\circ_8&:=O*\mathcal{S}_{K^\circ}(-d^\circ \cup \widetilde{g^\circ} \cup \widetilde{-i^\circ}), \end{aligned}\] which corresponds to the decomposition of \(K\) described before. The volume of \(K^\circ_i \ (i=1,2,\ldots,8)\) is denoted by \(|K^\circ_i|\). Denote by \(P_i \ (i=1,2,3)\) the orthogonal projections to the \(vw\)-plane, \(wu\)-plane and \(uv\)-plane, respectively. Since \(K^\circ \subset \R^3\) is strongly convex, the projections \(P_i(K^\circ) \ (i=1,2,3)\) of \(K^\circ\) are compact strongly convex domains in each plane.
## A volume estimate of a piece of \(K^\circ\) from below {#sec:3.3}
The next task is to estimate \(|K^\circ_i| \ (i=1,2,\ldots,8)\) from below. Consider \(K^\circ_1=O*\mathcal{S}_{K^\circ}(d^\circ \cup f^\circ \cup h^\circ)\), for instance. The boundary \(\partial K^\circ_1\) is the union of the cone \[O*(d^\circ \cup f^\circ \cup h^\circ)\] and \(K^\circ_1 \cap \partial K^\circ\). For any point \(P=(u,v,w) \in \R^3\), the signed volume of the cone \[P*(O*(d^\circ \cup f^\circ \cup h^\circ)) = \{ (1-\lambda)P+\lambda\] is described as \[\label{eq:b} \frac{1}{6} \int_0^1 \left( \begin{vmatrix} P \\ d^\circ(t) \\ {d^\circ}'(t) \end{vmatrix} + \begin{vmatrix} P \\ f^\circ(t) \\ {f^\circ}'(t) \end{vmatrix} + \begin{vmatrix} P \\ h^\circ(t) \\ {h^\circ}'(t) \end{vmatrix} \right)\,dt,\] where \(d^\circ(t) = (d^\circ_1(t), d^\circ_2(t), d^\circ_3(t))\) and \(|\cdot|\) denotes the determinant for a square matrix. Here we put \[\begin{aligned}
\overline{d^\circ_1}:= \int_0^1 \begin{vmatrix} d^\circ_2(t) & d^\circ_3(t) \\ {d^\circ_2}'(t) & {d_3^\circ}'(t) \end{vmatrix}\,dt, \overline{d^\circ_2}:= \int_0^1 \begin{vmatrix} d^\circ_3(t) & d^\circ_1(t) \\ {d^\circ_3}'(t) & {d^\circ_1}'(t) \end{vmatrix}\,dt, \overline{d^\circ_3}:= \int_0^1 \begin{vmatrix} d^\circ_1(t) & d^\circ_2(t) \\ {d^\circ_1}'(t) & {d^\circ_2}'(t) \end{vmatrix}\,dt, \end{aligned}\] then we have \[\int_0^1 \begin{vmatrix} P \\ d^\circ(t) \\ {d^\circ}'(t) \end{vmatrix}\,dt = (\overline{d^\circ_1}, \overline{d^\circ_2}, \overline{d^\circ_3}) \cdot (u,v,w).\] For simplicity, we put \(\overline{d^\circ} := (\overline{d^\circ_1}, \overline{d^\circ_2}, \overline{d^\circ_3})\). Similarly, we define \(\overline{e^\circ}, \overline{f^\circ}, \overline{g^\circ}, \overline{h^\circ}\), and \(\overline{i^\circ}\). Then the signed volume ([\[eq:b\]](#eq:b){reference-type="ref" reference="eq:b"}) of the cone \(P*(O*(d^\circ \cup f^\circ \cup h^\circ))\) is represented as \((\overline{d^\circ}+\overline{f^\circ}+\overline{h^\circ}) \cdot (u,v,w)/6\). If the cone \(K^\circ_1\) is convex in \(\R^3\) and \(P \in K^\circ_1\), then the quantity ([\[eq:b\]](#eq:b){reference-type="ref" reference="eq:b"}) coincides with the volume \(|P*(O*(d^\circ \cup f^\circ \cup h^\circ))|\) and it is less than \(|K^\circ_1|\). In case \(K^\circ_1 \subset \R^3\) is *not* convex or \(P \in K^\circ\) is *not* in \(K^\circ_1\), the comparison of the signed volume ([\[eq:b\]](#eq:b){reference-type="ref" reference="eq:b"}) with \(|K^\circ_1|\) is a nontrivial problem. However, under the condition that \(P\) lies in the convex set \(K^\circ\) the following holds, which is a key lemma to estimate \(\mathcal{P}(K)\).
The proof of Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"} is somewhat long, so we will prove it in the next subsection. To discuss other \(K^\circ_i\)'s we need further preparations. For the curve \(\widetilde{d^\circ}\) on \(\partial K^\circ\) defined by \(\widetilde{d^\circ}(t) = d^\circ(1-t) \ (0 \leq t \leq 1)\), we have \[\begin{aligned}
\overline{(\widetilde{d^\circ})} \cdot (u,v,w) =& \int_0^1 \begin{vmatrix} P \\ d^\circ(1-t) \\ (d^\circ(1-t))' \end{vmatrix}\,dt =- \int_0^1 \begin{vmatrix} P \\ d^\circ(1-t) \\ {d^\circ}'(1-t) \end{vmatrix}\,dt =-\overline{d^\circ} \cdot (u,v,w) \end{aligned}\] for any \(P=(u,v,w) \in \R^3\). Moreover, since \[\overline{(-d^\circ)} \cdot (u,v,w) = \int_0^1 \begin{vmatrix} P \\-d^\circ(t) \\-{d^\circ}'(t) \end{vmatrix}\,dt = \overline{d^\circ} \cdot (u,v,w),\] we obtain \[\overline{(\widetilde{d^\circ})} =-\overline{d^\circ}, \quad \overline{(-d^\circ)} = \overline{d^\circ}, \quad \overline{(-\widetilde{d^\circ})} = \overline{(\widetilde{-d^\circ})} =-\overline{d^\circ}.\] We can do the same observation for \(K^\circ_2\), \(K^\circ_3\), and \(K^\circ_4\). Since \(K^\circ\) is centrally symmetric, the volume of \(K^\circ_5\), \(K^\circ_6\), \(K^\circ_7\), and \(K^\circ_8\) equals that of \(K^\circ_3\), \(K^\circ_4\), \(K^\circ_1\), and \(K^\circ_2\), respectively. Summarizing the above arguments, we obtain
Furthermore, similar arguments for the convex body \(K\) give the following results. Note that the calculation is easier than the case of the polar \(K^\circ\), because \(K_i \ (i=1,2,\ldots,8)\) are convex.
In Section [3.5](#sec:3.5){reference-type="ref" reference="sec:3.5"}, these eight points \(R_i, S_i \ (i=1,2,3,4)\) will be effectively used as test points to estimate \(\mathcal{P}(K)\) from below.
## The proof of Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"} {#sec:3.4}
Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"} is a direct consequence of the following
The proof will involve a number of steps. First we consider the case of a "triangle" in the following sense. Let \(A_1, A_2, A_3 \in \partial K\) be mutually distinct points with \((A_1 \times A_2)\cdot A_3 >0\). For \((i,j)=(1,2),(2,3), (3,1)\), denote by \(C_{i,j}\) the oriented segment \(\mathcal{C}_K(A_i,A_j)\) on \(\partial K\), which is parametrized by \[l_{i,j}(t):=\rho_K((1-t) A_i + t A_j) \left((1-t) A_i + t A_j\right) \quad (0 \leq t \leq 1).\] Each \(C_{i,j}\) is also in the plane which through three points \(O\), \(A_i\), and \(A_j\). Let us consider the domain \(\mathcal{S}_K(c)\) on \(\partial K\) enclosed by \(c:=C_{1,2} \cup C_{2,3} \cup C_{3,1}\). We call this \(\mathcal{S}_K(c)\) a *triangle* on the smooth convex surface \(\partial K\). Putting \(S_{i,j} := O*C_{i,j}\) and \(D_{i,j} := P*S_{i,j}\), then \[\frac{1}{6} \int_{C_{i,j}} (P \times \vr) \cdot d \vr = \frac{1}{6} \int_0^1 (P \times l_{i,j}(t)) \cdot l_{i,j}'(t) \,dt = \frac{1}{6} \int_0^1 \begin{vmatrix} P \\ l_{i,j}(t) \\ l_{i,j}'(t) \end{vmatrix} \,dt\] is the signed volume of \(D_{i,j}\). Let \(|D_{i,j}|\) denote its absolute value. Note that all \(D_{i,j}\) mutually do not intersect their interiors. In this setting, the following lemma holds.
From Lemma [\[lem:14\]](#lem:14){reference-type="ref" reference="lem:14"}, we obtain the following
## A sharp estimate {#sec:3.5}
In this subsection, we give a sufficient condition of deducing inequality [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"} (see the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} below). In the following sections, for any \(K \in \hat{\K}\) we carefully select a linear transformation \(\mathcal{A}\) such that \(\mathcal{A}K \in \hat{\K}\) satisfies this condition. From the arguments in Section [3.3](#sec:3.3){reference-type="ref" reference="sec:3.3"}, we know that \(R_i \in K^\circ\) and \(S_i \in K\) \((i=1,2,3,4)\). By the definition of polar, we obtain \(R_i \cdot S_i \leq 1\). In other words, \[\label{eq:2} \begin{aligned} \left(\bar{d}+\bar{f}+\bar{h}\right)\cdot \left(\overline{d^\circ}+\overline{f^\circ}+\overline{h^\circ}\right) & \leq 36 |K_1| |K^\circ_1|, \\ \left(-\bar{d}+\bar{g}+\bar{i}\right)\cdot \left(-\overline{d^\circ}+\overline{g^\circ}+\overline{i^\circ}\right) & \leq 36 |K_2| |K^\circ_2|, \\ \left(-\bar{e}-\bar{g}+\bar{h}\right)\cdot \left(-\overline{e^\circ}-\overline{g^\circ}+\overline{h^\circ}\right) & \leq 36 |K_3| |K^\circ_3|, \\ \left(\bar{e}-\bar{f}+\bar{i}\right)\cdot \left(\overline{e^\circ}-\overline{f^\circ}+\overline{i^\circ}\right) & \leq 36 |K_4| |K^\circ_4|. \end{aligned}\] Here, assume that \[\label{eq:3} |K_1|=|K_2|=|K_3|=|K_4|, \quad \overline{d}= \overline{e}, \quad \overline{f}= \overline{g}, \quad \overline{h}= \overline{i}.\] We now check that this condition implies symmetric Mahler's conjecture for the three dimensional case. Indeed, the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} yields \(|K_i|=|K|/8\), and hence equations [\[eq:2\]](#eq:2){reference-type="eqref" reference="eq:2"} and [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} implies that \[\begin{aligned}
\left(\bar{d}+\bar{f}+\bar{h}\right)\cdot \left(\overline{d^\circ}+\overline{f^\circ}+\overline{h^\circ}\right) & \leq \frac{9}{2} \abs{K} \abs{K^\circ_1}, \\ \left(-\bar{d}+\bar{f}+\bar{h}\right)\cdot \left(-\overline{d^\circ}+\overline{g^\circ}+\overline{i^\circ}\right) & \leq \frac{9}{2} \abs{K} \abs{K^\circ_2}, \\ \left(-\bar{d}-\bar{f}+\bar{h}\right)\cdot \left(-\overline{e^\circ}-\overline{g^\circ}+\overline{h^\circ}\right) & \leq \frac{9}{2} \abs{K} \abs{K^\circ_3}, \\ \left(\bar{d}-\bar{f}+\bar{h}\right)\cdot \left(\overline{e^\circ}-\overline{f^\circ}+\overline{i^\circ}\right) & \leq \frac{9}{2} \abs{K} \abs{K^\circ_4}. \end{aligned}\] It then follows from these inequalities that \[\begin{aligned}
&\frac{9}{4} |K| |K^\circ| = \frac{9}{2} \abs{K} (|K^\circ_1|+|K^\circ_2|+|K^\circ_3|+|K^\circ_4|) \\ & \geq \overline{d} \cdot \left( \left(\overline{d^\circ}+\overline{f^\circ}+\overline{h^\circ}\right)-\left(-\overline{d^\circ}+\overline{g^\circ}+\overline{i^\circ}\right)-\left(-\overline{e^\circ}-\overline{g^\circ}+\overline{h^\circ}\right) +\left(\overline{e^\circ}-\overline{f^\circ}+\overline{i^\circ}\right) \right) \\ &\quad + \overline{f} \cdot \left( \left(\overline{d^\circ}+\overline{f^\circ}+\overline{h^\circ}\right) +\left(-\overline{d^\circ}+\overline{g^\circ}+\overline{i^\circ}\right)-\left(-\overline{e^\circ}-\overline{g^\circ}+\overline{h^\circ}\right)-\left(\overline{e^\circ}-\overline{f^\circ}+\overline{i^\circ}\right) \right) \\ &\quad + \overline{h} \cdot \left( \left(\overline{d^\circ}+\overline{f^\circ}+\overline{h^\circ}\right) +\left(-\overline{d^\circ}+\overline{g^\circ}+\overline{i^\circ}\right) +\left(-\overline{e^\circ}-\overline{g^\circ}+\overline{h^\circ}\right) +\left(\overline{e^\circ}-\overline{f^\circ}+\overline{i^\circ}\right) \right) \\ &= 2 \overline{d} \cdot \left(\overline{d^\circ} + \overline{e^\circ}\right) + 2 \overline{f} \cdot \left(\overline{f^\circ} + \overline{g^\circ}\right) + 2 \overline{h} \cdot \left(\overline{h^\circ} + \overline{i^\circ}\right). \end{aligned}\] Let us examine the first term in the last line. Recalling the definition of \(\overline{d^\circ}=(\overline{d^\circ_1}, \overline{d^\circ_2}, \overline{d^\circ_3})\), its \(u\)-component \(\overline{d^\circ_1}\) is twice the area of the convex region \(O*P_1(d^\circ)\) in the \(vw\)-plane. Similarly, \(\overline{e^\circ_1}\) is twice \(|O*P_1(e^\circ)|\). Since \(K^\circ \in \hatK\), the convex domain \(P_1(K^\circ)\) in the \(vw\)-plane is also centrally symmetric, and hence \[\overline{d^\circ_1}+\overline{e^\circ_1} = 2(|O*P_1(d^\circ)|+|O*P_1(e^\circ)|) = |P_1(K^\circ)|.\] Again, since the curve \(d \subset K\) lies in the \(yz\)-plane, we have \(\overline{d} = (\overline{d_1},0,0)\), where \[\overline{d_1} = \int_0^1 \begin{vmatrix} d_2(t) & d_3(t) \\ d_2'(t) & d_3'(t) \end{vmatrix} \,dt.\] Here \(\overline{d_1}\) is nothing but twice the area of the convex region \(O*d\) in the \(yz\)-plane. From the assumption that \(\overline{d}=\overline{e}\) in the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}, we have \(\overline{d_1}=|Q_1(K)|/2\). Therefore, the first term \(2 \overline{d} \cdot \left(\overline{d^\circ} + \overline{e^\circ}\right)\) is equal to \(|Q_1(K)|\,|P_1(K^\circ)|\). Applying the same argument to the second and the third terms, we have \[\overline{f^\circ_2}+\overline{g^\circ_2} = |P_2(K^\circ)|, \quad \overline{h^\circ_3}+\overline{i^\circ_3} = |P_3(K^\circ)|\] and \[2 \overline{f} = (0,|Q_2(K)|,0), \quad 2 \overline{h} = (0,0,|Q_3(K)|).\] Consequently, we obtain \[\frac{9}{4} |K| |K^\circ| \geq |Q_1(K)|\, |P_1(K^\circ)|+ |Q_2(K)|\, |P_2(K^\circ)|+ |Q_3(K)|\, |P_3(K^\circ)|.\] It is well-known that convex domains \(Q_i(K)\) and \(P_i(K^\circ)\) are polar bodies of each other (e.g. \*p. 274). By Mahler's theorem (see also, ,, and Section [2](#sec:2){reference-type="ref" reference="sec:2"}), we have \(|Q_i(K)|\, |P_i(K^\circ)| \geq 4^2/2! =8 \ (i=1,2,3)\). Thus \[|K| \, |K^\circ| \geq \frac{4}{9} \times 8 \times 3 = \frac{32}{3} =\frac{4^3}{3!}\] holds. We conclude that the Mahler conjecture is true for \(K \in \hat{\K}\) *satisfying the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}*.
# The case with a hyperplane symmetry {#sec:4}
From the conclusion of the previous section, it is important to find out when the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} holds. We start with the following simple observation.
From now on, suppose that \(K \in \hat{\K}\) is symmetric with respect to a plane. Since the volume product is invariant under linear transformations of \(K\), we may assume that the hyperplane of symmetry is the \(yz\)-plane. In this case, the area of \(O*f\) and that of \(O*(-g)\) coincide. Taking the signs into consideration, \(\overline{f}=\overline{g}\) holds. Similarly, \(\overline{h}=\overline{i}\) holds. Concerning the volume of \(K_i \ (i=1,\ldots,8)\), we have \[|K_1|= |K_2|= |K_8|, \quad |K_4|= |K_3|= |K_5|.\] Consequently, the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} holds, provided \(K\) satisfies \[\label{eq:4} \overline{d}=\overline{e}, \quad |K_1|= |K_4|.\] For a given convex body \(K \in \hat{\K}\), we consider the image of \(K\) by the following linear transformation: \[A(\alpha) X(\theta) \begin{pmatrix} x \\ y \\ z \end{pmatrix} = \begin{pmatrix} 1 & 0 & 0 \\ 0 & 1 & \alpha \\ 0 & 0 & 1 \end{pmatrix} \begin{pmatrix} 1 & 0 & 0 \\ 0 & \cos \theta &-\sin \theta \\ 0 & \sin \theta & \cos \theta \end{pmatrix} \begin{pmatrix} x \\ y \\ z \end{pmatrix}.\] For \(K \in \hat{\K}\), we introduced vectors \(\overline{d}=(d_1,d_2,d_3)\), etc., in Section [3.3](#sec:3.3){reference-type="ref" reference="sec:3.3"}. From now on, to clarify the dependence on the convex body, we denote the vector \(\overline{d}\) for \(A(\alpha)X(\theta)K \in \hat{\K}\) by \(\overline{d}(A(\alpha)X(\theta)K)\), etc. Similarly, for the image \(A(\alpha)X(\theta)K \in \hat{\K}\) of a convex body \(K\), we denote by \((A(\alpha)X(\theta)K)_i\) \((i=1, \dots, 8)\) the decomposition in Section [3.2](#sec:3.2){reference-type="ref" reference="sec:3.2"}. First, since \(A(\alpha)\) and \(X(\theta)\) keep the symmetry of \(K\) with respect to the \(yz\)-plane, we have \[\begin{aligned}
&\overline{f}(A(\alpha)X(\theta)K) = \overline{g}(A(\alpha)X(\theta)K), \\ &\overline{h}(A(\alpha)X(\theta)K) = \overline{i}(A(\alpha)X(\theta)K), \\ &|(A(\alpha)X(\theta)K)_1|=|(A(\alpha)X(\theta)K)_2|=|(A(\alpha)X(\theta)K)_8|, \\ &|(A(\alpha)X(\theta)K)_4|=|(A(\alpha)X(\theta)K)_3|=|(A(\alpha)X(\theta)K)_5| \end{aligned}\] for any \(\alpha, \theta\) \(\in \mathbb R\). Next, for a given \(\theta \in \mathbb R\) we can choose \(\alpha=\alpha(\theta)\) \(\in \R\) to satisfy \[|(A(\alpha)X(\theta)K)_1|= |(A(\alpha)X(\theta)K)_4|,\] see Proposition [\[prop:3\]](#prop:3){reference-type="ref" reference="prop:3"} below in the special case that \(\Phi(A(\alpha)X(\theta)K)=\Psi(A(\alpha)X(\theta)K)=\pi/2\), \(|\Delta_1(A(\alpha)X(\theta)K)|=|\Delta_2(A(\alpha)X(\theta)K)|\), \(|\Delta_3(A(\alpha)X(\theta)K)|=|\Delta_4(A(\alpha)X(\theta)K)|\) for \(X(\theta)K\). Note that such \(\alpha\) is uniquely determined for each \(\theta\), and \(\alpha\) depends on \(\theta\) continuously. Moreover, we can easily see that there exists some \(\theta_0\) such that \(\alpha(\theta_0)=0\). In this setting, we obtain \(\alpha(\theta_0+\pi/2)=0\) and \[\label{eq:20} \begin{aligned} \overline{d}(A(\alpha(\theta_0+\pi/2))X(\theta_0+\pi/2)K) &= \overline{d}(X(\theta_0+\pi/2)K) = \overline{e}(A(\alpha(\theta_0))X(\theta_0)K), \\ \overline{e}(A(\alpha(\theta_0+\pi/2))X(\theta_0+\pi/2)K) &= \overline{e}(X(\theta_0+\pi/2)K) = \overline{d}(A(\alpha(\theta_0))X(\theta_0)K). \end{aligned}\] Next, we put \[F(\theta):= \overline{d} (A(\alpha(\theta))X(\theta)K)- \overline{e} (A(\alpha(\theta))X(\theta)K).\] Since \(\alpha\) is continuous with respect to \(\theta\), \(F\) is continuous on \(\R\). By [\[eq:20\]](#eq:20){reference-type="eqref" reference="eq:20"}, we have \(F(\theta_0+\pi/2)=-F(\theta_0)\). By using the intermediate value theorem, there exists \(\theta_1 \in [\theta_0, \theta_0+\pi/2]\) such that \(F(\theta_1)=0\), which yields \[\overline{d} (A(\alpha(\theta_1))X(\theta_1)K) = \overline{e} (A(\alpha(\theta_1))X(\theta_1)K).\] Consequently, \(A(\alpha(\theta_1))X(\theta_1)K \in \hatK\) satisfies the condition [\[eq:4\]](#eq:4){reference-type="eqref" reference="eq:4"}, and hence [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}. Since \(A(\alpha(\theta_1))X(\theta_1)\) is a linear transformation of \(\R^3\), we obtain \[\mathcal{P}(K) = \mathcal{P}(A(\alpha(\theta_1))X(\theta_1)K) \geq \frac{32}{3}.\] Thus, we have proved the following
# The general case {#sec:5}
## Notations {#sec:5.1}
In order to prove [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"} for general \(K \in \K^3_0\), we first consider some linear transform of \(K \in \hat{\mathcal K}\). For this purpose, we introduce some notations. For \(\theta, \phi, \psi \in \R\), \(X(\theta)\), \(Y(\phi)\), \(Z(\psi)\) denote rotations in \(\R^3\) as follows: \[\begin{aligned}
X(\theta) &= \begin{pmatrix} 1 & 0 & 0 \\ 0 & \cos \theta &-\sin \theta \\ 0 & \sin \theta & \cos \theta \end{pmatrix}, \quad Y(\phi)= \begin{pmatrix} \cos \phi & 0 & \sin \phi \\ 0 & 1 & 0 \\-\sin \phi & 0 & \cos \phi \end{pmatrix}, \\ Z(\psi) &= \begin{pmatrix} \cos \psi &-\sin \psi & 0 \\ \sin \psi & \cos \psi & 0\\ 0 & 0 & 1 \end{pmatrix}. \end{aligned}\] Hereafter, we put \(K(\theta, \phi, \psi):=X(\theta)Y(\phi)Z(\psi)K\). Each \(\Delta_i(K)\) denotes the part of \(K\) included in the \(i\)-th octant: \[\begin{aligned}
\Delta_1(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \geq 0, y \geq 0, z \geq 0 \right\}, \\ \Delta_2(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \leq 0, y \geq 0, z \geq 0 \right\}, \\ \Delta_3(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \leq 0, y \leq 0, z \geq 0 \right\}, \\ \Delta_4(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \geq 0, y \leq 0, z \geq 0 \right\}, \\ \Delta_5(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \geq 0, y \geq 0, z \leq 0 \right\}, \\ \Delta_6(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \leq 0, y \geq 0, z \leq 0 \right\}, \\ \Delta_7(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \leq 0, y \leq 0, z \leq 0 \right\}, \\ \Delta_8(K) &:= K \cap \left\{ (x,y,z) \in \R^3; x \geq 0, y \leq 0, z \leq 0 \right\}. \end{aligned}\] As in Section [3.2](#sec:3.2){reference-type="ref" reference="sec:3.2"}, we put \[A(K):=\rho_K((1,0,0)) (1,0,0), B(K):=\rho_K((0,1,0)) (0,1,0), C(K):=\rho_K((0,0,1)) (0,0,1)\] and \[\begin{aligned}
d(K)&:=\mathcal{C}_{K}(B(K), C(K)), & e(K)&:=\mathcal{C}_{K}(C(K),-B(K)), \\ f(K)&:=\mathcal{C}_{K}(C(K), A(K)), & g(K)&:=\mathcal{C}_{K}(A(K),-C(K)), \\ h(K)&:=\mathcal{C}_{K}(A(K), B(K)), & i(K)&:=\mathcal{C}_{K}(B(K),-A(K)). \end{aligned}\] Then we see that \(\bar{d}(K)\) is calculated as \[\begin{aligned}
& \left( \int_0^1 \begin{vmatrix} d_2(K)(t) & d_3(K)(t) \\ d_2(K)'(t) & d_3(K)'(t) \end{vmatrix}\,dt, \int_0^1 \begin{vmatrix} d_3(K)(t) & d_1(K)(t) \\ d_3(K)'(t) & d_1(K)'(t) \end{vmatrix}\,dt, \int_0^1 \begin{vmatrix} d_1(K)(t) & d_2(K)(t) \\ d_1(K)'(t) & d_2(K)'(t) \end{vmatrix}\,dt \right)\\ &= 2 \left( |P_1(O*d(K))|, |P_2(O*d(K))|, |P_3(O*d(K))| \right) \\ &= 2 \left(|O*d(K)|, 0,0\right), \end{aligned}\] where \(P_1, P_2\), and \(P_3\) denote projections to the \(yz\)-plane, \(zx\)-plane, and \(xy\)-plane, respectively. Similarly, we get \[\begin{aligned}
\bar{d}(K) &= 2 \left(|O*d(K)|,0,0\right), & \bar{e}(K) &= 2 \left(|O*e(K)|,0,0\right), \\ \bar{f}(K) &= 2 \left(0,|O*f(K)|,0\right), & \bar{g}(K) &= 2 \left(0,|O*g(K)|,0\right), \\ \bar{h}(K) &= 2 \left(0,0,|O*h(K)|\right), & \bar{i}(K) &= 2 \left(0,0,|O*i(K)|\right). \end{aligned}\] Note that \[\begin{aligned}
\Delta_1(K)&=O*\mathcal{S}_K(d(K) \cup f(K) \cup h(K)), & \Delta_2(K)&=O*\mathcal{S}_K(\widetilde{d(K)} \cup i(K) \cup-g(K)), \\ \Delta_3(K)&=O*\mathcal{S}_K(\widetilde{e(K)} \cup \widetilde{-g(K)} \cup-h(K)), & \Delta_4(K)&=O*\mathcal{S}_K(e(K) \cup-i(K) \cup \widetilde{f(K)}) \end{aligned}\] and \[\begin{aligned}
O*d(K)&= K \cap \left\{ (x,y,z) \in \R^3; x = 0, y \geq 0, z \geq 0 \right\}, \\ O*e(K)&= K \cap \left\{ (x,y,z) \in \R^3; x = 0, y \leq 0, z \geq 0 \right\}, \\ O*f(K)&= K \cap \left\{ (x,y,z) \in \R^3; x \geq 0, y = 0, z \geq 0 \right\}, \\ O*g(K)&= K \cap \left\{ (x,y,z) \in \R^3; x \geq 0, y = 0, z \leq 0 \right\}, \\ O*h(K)&= K \cap \left\{ (x,y,z) \in \R^3; x \geq 0, y \geq 0, z = 0 \right\}, \\ O*i(K)&= K \cap \left\{ (x,y,z) \in \R^3; x \leq 0, y \geq 0, z = 0 \right\}. \end{aligned}\] Therefore, the condition [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} is equivalent to the following equations: \[\label{eq:23} \begin{cases} |\Delta_1(K)|= |\Delta_2(K)|= |\Delta_3(K)|= |\Delta_4(K)|, \\ |O*d(K)|= |O*e(K)|, \\ |O*f(K)|= |O*g(K)|, \\ |O*h(K)|= |O*i(K)|. \end{cases}\] According to the discussion in Section [3.5](#sec:3.5){reference-type="ref" reference="sec:3.5"}, we obtain
## Linear transform {#sec:5.2}
In this section, we choose a linear transform \(\mathcal{A}=\mathcal{A}(K)\) (depends on \(K\)) which satisfies \[\label{eq:22} \begin{aligned} |O*f(\mathcal{A}K)| = |O*g(\mathcal{A}K)|, &\quad |O*h(\mathcal{A}K)| = |O*i(\mathcal{A}K)|, \\ |\Delta_1(\mathcal{A}K)|+|\Delta_2(\mathcal{A}K)| &= |\Delta_3(\mathcal{A}K)|+|\Delta_4(\mathcal{A}K)|, \end{aligned}\] where \(\mathcal{A}K=\mathcal{A}(K)K\). For the purpose, we use the following spherical coordinates: \[P(\alpha, \beta) =(\cos \alpha, \sin \alpha \cos \beta, \sin \alpha \sin \beta) \in S^2.\] First, we choose the angle \(\Theta(K) \in (0,\pi)\) which satisfies that \[\label{eq:21} \int_0^{\Theta(K)} d \beta \int_0^{\pi} \rho_{K}^3(P(\alpha,\beta)) \sin \alpha \,d \alpha = \int_{\Theta(K)}^\pi d \beta \int_0^{\pi} \rho_{K}^3(P(\alpha,\beta)) \sin \alpha \,d \alpha.\] Since the left-hand side is increasing and the right-hand side is decreasing with respect to \(\Theta(K)\), we see that \(\Theta(K)\) is uniquely determined. Actually, we have the following
Next, we introduce the quantities \(\Phi(K) \in (0,\pi)\) and \(\Psi(K) \in (0,\pi)\) defined by \[\begin{aligned}
\int_0^{\Phi(K)} \rho_{K}^2(P(\alpha,0)) \,d\alpha &= \int_{\Phi(K)}^\pi \rho_{K}^2(P(\alpha,0)) \,d\alpha, \\ \int_0^{\Psi(K)} \rho_{K}^2(P(\alpha,\Theta(K))) \,d\alpha &= \int_{\Psi(K)}^\pi \rho_{K}^2(P(\alpha,\Theta(K))) \,d\alpha. \end{aligned}\] Similarly as \(\Theta(K)\), the above \(\Phi(K)\) and \(\Psi(K)\) are uniquely determined for each \(K \in \hatK\), and \(\Phi(\theta,\phi,\psi) :=\Phi(K(\theta,\phi,\psi))\) and \(\Psi(\theta,\phi,\psi) :=\Psi(K(\theta,\phi,\psi))\) are smooth functions on \(\R^3\). Now, we define \(\mathcal{A}=\mathcal{A}(K)\) as \[\begin{aligned}
\linearop(K)&:= \begin{pmatrix} 1 & \frac{1}{\tan (\Phi(K))} & \frac{1}{\sin(\Theta(K)) \tan (\Psi(K))} \\ 0 & 1 & \frac{1}{\tan(\Theta(K))} \\ 0 & 0 & 1 \end{pmatrix}^{-1}. \end{aligned}\]
## Definition of a smooth map \((F,G,H)\) {#sec:5.3}
For simplicity, we denotes \(\linearop K:=\mathcal{A}(K) K\) and \[\linearop K(\theta, \phi, \psi):= \mathcal{A}(K(\theta, \phi, \psi))K(\theta, \phi, \psi) = \mathcal{A}(X(\theta)Y(\phi)Z(\psi)K)X(\theta)Y(\phi)Z(\psi)K.\] For every \(K \in \hatK\), we define real numbers \(F, G, H\) by \[\begin{aligned}
F(K) &:= |O*d(\linearop K)|-|O*e(\linearop K)|, \\ G(K) &:= |\Delta_1(\linearop K)|+ |\Delta_3(\linearop K)|-|\Delta_2(\linearop K)|-|\Delta_4(\linearop K)|, \\ H(K) &:= |\Delta_1(\linearop K)|+ |\Delta_4(\linearop K)|-|\Delta_2(\linearop K)|-|\Delta_3(\linearop K)|, \end{aligned}\] and introduce the following functions on \(\R^3\): \[\begin{aligned}
F(\theta, \phi, \psi)&:= F({\linearop K}(\theta, \phi, \psi)), \\ G(\theta, \phi, \psi)&:= G({\linearop K}(\theta, \phi, \psi)), \\ H(\theta, \phi, \psi)&:= H({\linearop K}(\theta, \phi, \psi)). \end{aligned}\] If we can find a zero \((\theta, \phi, \psi)\) of \((F,G,H)\), then \({\linearop K}(\theta, \phi, \psi)\) satisfies the condition [\[eq:23\]](#eq:23){reference-type="eqref" reference="eq:23"} by Proposition [\[prop:3\]](#prop:3){reference-type="ref" reference="prop:3"}. To show the existence of such a zero, we consider the following region in \(\R^3\): \[D = \left\{(\theta, \phi, \psi) \in \R^3; 0 \leq \phi \leq \pi, 0 \leq \psi \leq \pi, 0 \leq \theta \leq \pi-\Theta(0,\phi,\psi) \right\}.\] Note that \(\Theta\) is a smooth function introduced in the previous subsection and \(0 < \Theta(0,\phi,\psi) < \pi\). We divide the boundary \(\partial D\) into the following six parts: \[\begin{aligned}
M_1 &:= \left\{ (0, \phi, \psi) \in \R^3; 0 \leq \phi \leq \pi, 0 \leq \psi \leq \pi\right\}, \\ M_2 &:= \left\{ (\pi-\Theta(0,\phi,\psi), \phi, \psi) \in \R^3; 0 \leq \phi \leq \pi, 0 \leq \psi \leq \pi\right\}, \\ M_3 &:= \left\{(\theta, 0, \psi) \in \R^3; 0 \leq \psi \leq \pi, 0 \leq \theta \leq \pi-\Theta(0,0,\psi) \right\}, \\ M_4 &:= \left\{(\theta, \pi, \psi) \in \R^3; 0 \leq \psi \leq \pi, 0 \leq \theta \leq \pi-\Theta(0,\pi,\psi) \right\}, \\ M_5 &:= \left\{(\theta, \phi, 0) \in \R^3; 0 \leq \phi \leq \pi, 0 \leq \theta \leq \pi-\Theta(0,\phi,0) \right\}, \\ M_6 &:= \left\{(\theta, \phi, \pi) \in \R^3; 0 \leq \phi \leq \pi, 0 \leq \theta \leq \pi-\Theta(0,\phi,\pi) \right\}. \end{aligned}\] Then, we have \[\partial D = M_1 \cup \dots \cup M_6\] and \(M_i \cap M_j\) is empty or a curve for each \(i \not= j\). In the case where \((F,G,H)\) has no zeros on \(\partial D\), we can use the degree of a map \[\mathscr{F}:=\frac{(F,G,H)}{\sqrt{F^2+G^2+H^2}}: \partial D \rightarrow S^2\] to find a zero in the interior of \(D\). In order to calculate the degree of \(\mathscr{F}\), in the rest of this section, we show the identities [\[eq:8\]](#eq:8){reference-type="eqref" reference="eq:8"}, [\[eq:10\]](#eq:10){reference-type="eqref" reference="eq:10"}, and [\[eq:11\]](#eq:11){reference-type="eqref" reference="eq:11"} in Proposition [\[prop:7\]](#prop:7){reference-type="ref" reference="prop:7"} below.
## Some formulas {#sec:5.4}
We prepare some formulas.
## Identities relating to the rotation \(X(\pi-\Theta(K))\) and \(X(\pi)\)
Here we examine the behavior of the functions \(F\), \(G\), and \(H\) under the rotation \(X\) of \(K\).
Using the above lemmas, we get identities relating to the rotation \(X(\pi)\).
## Identities relating to the rotation \(Y(\pi)\)
## Identities relating to the rotation \(Z(\pi)\)
## Key identities on \(\partial D\) {#sec:5.8}
By using the lemmas prepared in the above subsections, we show Proposition [\[prop:7\]](#prop:7){reference-type="ref" reference="prop:7"} below, which states key identities on \(\partial D\) for the degree calculation in the next section. To begin with, we introduce the map \(\Gamma\) defined by \[\Gamma_\psi(\theta):=\pi-\Theta(\theta,0,\psi)+\theta.\]
# Calculation of degree {#sec:6}
## Setting
We use new coordinates \((s, \phi, \psi)\) instead of \((\theta, \phi, \psi)\), where \[s= \frac{\theta}{\pi-\Theta(0,\phi,\psi)}.\] For the new coordinates \((s, \phi, \psi)\), we see that \[\begin{aligned}
M_1 &= \{0\} \times [0,\pi]\times[0,\pi], & M_3 &= [0,1]\times\{0\}\times[0,\pi], & M_5 &= [0,1]\times[0,\pi]\times\{0\}, \\ M_2 &= \{1\} \times [0,\pi]\times[0,\pi], & M_4 &= [0,1]\times\{\pi\}\times[0,\pi], & M_6 &= [0,1]\times[0,\pi]\times\{\pi\}. \end{aligned}\] We define \[\begin{aligned} \hat{F}(s, \phi, \psi) &:= F((\pi-\Theta(0,\phi,\psi))s, \phi, \psi), \\ \hat{G}(s, \phi, \psi) &:= G((\pi-\Theta(0,\phi,\psi))s, \phi, \psi), \\ \hat{H}(s, \phi, \psi) &:= H((\pi-\Theta(0,\phi,\psi))s, \phi, \psi). \end{aligned}\] Note that \[D= \left\{(s, \phi, \psi) \in \R^3; 0 \leq s \leq 1, 0 \leq \phi \leq \pi, 0 \leq \psi \leq \pi\right\}\] and \((\hat{F},\hat{G},\hat{H}) \in C^\infty(D, \R^3)\). Then \((\hat{F},\hat{G},\hat{H})|_{\partial D}\) is a triplet of continuous functions on \(\partial D=M_1 \cup \dots \cup M_6\), of class \(C^{\infty}\) in the interior of each \(M_i\) for \(i=1,\dots,6\). Hence, \((\hat{F},\hat{G},\hat{H})|_{\partial D}\) can be regarded as a piecewise smooth vector field on \(\partial D\). It follows from Proposition [\[prop:7\]](#prop:7){reference-type="ref" reference="prop:7"}, [\[eq:44\]](#eq:44){reference-type="eqref" reference="eq:44"}, and [\[eq:45\]](#eq:45){reference-type="eqref" reference="eq:45"} that \((\hat{F},\hat{G},\hat{H})\) satisfies the identities: \[\label{eq:12} \begin{aligned} \hat{F}(1,\phi,\psi)&=-\hat{F}(0,\phi,\psi), \\ \hat{G}(1,\phi,\psi)&=-\hat{H}(0,\phi,\psi), \\ \hat{H}(1,\phi,\psi)&=\hat{G}(0,\phi,\psi) \end{aligned}\] for \(\phi, \psi \in [0,\pi]\) and \[\label{eq:13} \begin{aligned} \hat{F}(s,\pi,\psi) &= \hat{F}(T_\psi(s),0,\psi), \\ \hat{G}(s,\pi,\psi) &=-\hat{H}(T_\psi(s),0,\psi), \\ \hat{H}(s,\pi,\psi) &=-\hat{G}(T_\psi(s),0,\psi) \end{aligned}\] for \(s \in [0,1], \psi \in [0,\pi]\) and \[\label{eq:14} \begin{aligned} \hat{F}(s,\phi,\pi) &= \hat{F}(s,\pi-\phi,0), \\ \hat{G}(s,\phi,\pi) &=-\hat{G}(s,\pi-\phi,0), \\ \hat{H}(s,\phi,\pi) &=-\hat{H}(s,\pi-\phi,0) \end{aligned}\] for \(s \in [0,1], \phi \in [0,\pi]\), where \(T_\psi:[0,1] \to [0,1]\) is defined by \[T_\psi(s) := \frac{\Gamma_\psi^{-1}\left(\pi-\Theta(0,0,\psi)s\right)}{\pi-\Theta(0,0,\psi)}.\] By Lemma [\[lem:11\]](#lem:11){reference-type="ref" reference="lem:11"}, \(T_\psi(s)\) is smooth with respect to \(\psi\) and \(s\), decreasing with respect to \(s\), \(T_\psi(0)=1\), and \(T_\psi(1)=0\). As we mentioned in Section [5.3](#sec:5.3){reference-type="ref" reference="sec:5.3"}, if we can find a zero \((\theta,\phi,\psi)\) in \(D\) of the vector field \((F,G,H)\), then the equations [\[eq:23\]](#eq:23){reference-type="eqref" reference="eq:23"} in Section [5.1](#sec:5.1){reference-type="ref" reference="sec:5.1"} are satisfied and the proof of 3-dimensional symmetric Mahler conjecture is completed. In the following two subsections, we consider the case that \((F,G,H)\) has no zeros on the boundary \(\partial D\). Now, assume that \((F,G,H) \not=0\) on \(\partial D\), that is, \((\hat{F},\hat{G},\hat{H}) \not=0\) on \(\partial D\). Then we shall calculate the degree of \[\mathscr{F}:=\frac{(\hat{F},\hat{G},\hat{H})}{\sqrt{\hat{F}^2+\hat{G}^2+\hat{H}^2}}:\partial D \to S^2.\]
## Reduction {#sec:6.2}
For notational convenience, in Sections [6.2](#sec:6.2){reference-type="ref" reference="sec:6.2"}, [6.3](#sec:6.3){reference-type="ref" reference="sec:6.3"}, and Appendix [\[sec:A\]](#sec:A){reference-type="ref" reference="sec:A"}, we denote the above \(\hat{F}\), \(\hat{G}\), and \(\hat{H}\) by \(F\), \(G\), and \(H\), respectively, and consider \[\mathscr{F}=\frac{(F,G,H)}{\sqrt{F^2+G^2+H^2}}:\partial D \to S^2.\] In order to calculate \(\deg \mathscr{F}\), we first need to modify the map \(\mathscr{F}\) so as to satisfy the property that \((\pm 1,0,0)\) are its regular values.
We give a proof of Proposition [\[prop:4\]](#prop:4){reference-type="ref" reference="prop:4"} in Appendix [\[sec:A\]](#sec:A){reference-type="ref" reference="sec:A"}. Owing to Proposition [\[prop:4\]](#prop:4){reference-type="ref" reference="prop:4"}, in what follows we may assume that \[\label{eq:46} \mathscr{F}(\partial M_i) \not\ni (\pm 1,0,0) \text{ for } i=1,\dots,6,\] and \((\pm 1,0,0)\) are regular values of \(\mathscr{F}\). Note that \(\mathscr{F}(s, \phi, \psi)=(\pm 1,0,0)\) if and only if \(G(s, \phi, \psi)=H(s, \phi, \psi)=0\) and \(\sgn F(s, \phi, \psi)=\pm 1\). Hence, the inverse image of \((\pm 1,0,0)\) by \(\mathscr{F} \rvert_{M_1}\) corresponds to the zeros of the vector field \((G,H)\) on \(M_1\). By the condition [\[eq:46\]](#eq:46){reference-type="eqref" reference="eq:46"}, \((G,H) \not= (0,0)\) on \(\partial M_1\) and the map \[\mathscr{G}:= \frac{(G,H)}{\sqrt{G^2+H^2}}: \partial M_1 \rightarrow S^1\] is well-defined.
## Calculation of \(w(\mathscr{G},(0,0))\) {#sec:6.3}
Proposition [\[prop:5\]](#prop:5){reference-type="ref" reference="prop:5"} reduces the problem to the following
## Proof of the main theorem (inequality part)
Let us prove the inequality part of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}. Note that a convex body \(K \in \K_0^3\) can be approximated by an element of \(\hatK\). The following is a direct consequence of a general approximation procedure for convex bodies by R. Schneider.
Since the volume is continuous with respect to the Hausdorff distance, it suffices to show the inequality \(\eqref{eq:1}\) for \(K \in \hatK\). Fix \(K \in \hatK\). We show the following
We denote the zero of \((\hat{F},\hat{G},\hat{H})\) by \((s_0,\phi_0,\psi_0)\). Putting \(\theta_0=(\pi-\Theta(0,\phi_0,\psi_0))s_0\), it means that \((F,G,H)=(0,0,0)\) at \((\theta_0,\phi_0,\psi_0)\). Thus the convex body \[\linearop K(\theta_0,\phi_0,\psi_0)=\mathcal{A}(X(\theta_0) Y(\phi_0) Z(\psi_0)K) X(\theta_0) Y(\phi_0) Z(\psi_0)K\] satisfies the key equality [\[eq:23\]](#eq:23){reference-type="eqref" reference="eq:23"}. Since \(\mathcal{A}(X(\theta_0) Y(\phi_0) Z(\psi_0)K) X(\theta_0) Y(\phi_0) Z(\psi_0)\) is a linear transformation, by Proposition [\[prop:2\]](#prop:2){reference-type="ref" reference="prop:2"} we obtain \[\mathcal{P}(K) = \mathcal{P}\left( \linearop K(\theta_0,\phi_0,\psi_0) \right) \geq \frac{32}{3}.\] Consequently, every \(K \in \hatK\) satisfies [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"}. Thus we complete the proof of the Mahler conjecture for any centrally symmetric convex body \(K \in \K_0^3\).
# The equality case {#sec:7}
In this section, we show the equality part of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}. That is,
## Dual face {#sec:7.1}
Let \(K \in \K^n_0\) be a centrally symmetric convex polytope in \(\R^n\). Then \(K^\circ\) is also a centrally symmetric convex polytope. For a face \(F\) of \(K\), its *dual face* \(F^\circ \subset K^\circ\) is defined by \[F^\circ:= \left\{ Q \in K^\circ; P \cdot Q=1 \text{ for any } P \in F \right\}.\] For a \(k\)-dimensional face \(F\), the dimension of \(F^\circ\) is \(n-k-1\). Assume \(K \in \K^2_0\) is a polygon. If a face \(F \subset K\) is a line segment which contains two points \((a,b)\) and \((c,d)\), then its dual face \(F^\circ\) is a vertex of \(K^\circ\). Putting \(F^\circ=\{(\alpha,\beta)\}\), by the definition of \(F^\circ\), we obtain that \[\begin{pmatrix} a & b \\ c & d \end{pmatrix} \begin{pmatrix} \alpha \\ \beta \end{pmatrix} = \begin{pmatrix} 1 \\ 1 \end{pmatrix}.\] Hence, \[\begin{pmatrix} \alpha \\ \beta \end{pmatrix} = \begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} \begin{pmatrix} 1 \\ 1 \end{pmatrix} = \frac{1}{ad-bc} \begin{pmatrix} d-b \\ a-c \end{pmatrix}.\] Assume \(K \in \K^3_0\) is a polytope. If a \(2\)-dimensional face \(F\) of \(K\) contains three points \((x_i, y_i, z_i)\) \((i=1,2,3)\) which do not on the same straight line, then its dual face \(F^\circ\) is a vertex of \(K^\circ\). Putting \(F^\circ=\{(u,v,w)\}\), we similarly obtain \[\begin{pmatrix} x_1 & y_1 & z_1 \\ x_2 & y_2 & z_2 \\ x_3 & y_3 & z_3 \end{pmatrix} \begin{pmatrix} u \\ v \\ w \end{pmatrix} = \begin{pmatrix} 1 \\ 1 \\ 1 \end{pmatrix}.\] Hence, \[\begin{pmatrix} u \\ v \\ w \end{pmatrix} = \begin{pmatrix} x_1 & y_1 & z_1 \\ x_2 & y_2 & z_2 \\ x_3 & y_3 & z_3 \end{pmatrix}^{-1} \begin{pmatrix} 1 \\ 1 \\ 1 \end{pmatrix}.\]
## Inequality
For a general convex body \(K\), we need the inequality similarly as Lemma [\[lem:14\]](#lem:14){reference-type="ref" reference="lem:14"}. As in Section [3.4](#sec:3.4){reference-type="ref" reference="sec:3.4"}, let \(A_i\) \((i=1,2,3)\) be points with \(A_1 \cdot (A_2 \times A_3)>0\), and consider \(C_{i,j}=\mathcal{C}_K(A_i, A_j)\) for \((i,j)=(1,2),(2,3),(3,1)\).
## Linear transformation of \(K \in \K_0^3\)
To prove Theorem [\[prop:14\]](#prop:14){reference-type="ref" reference="prop:14"}, for a general centrally symmetric convex body \(K\), we need to find a transformation \(\mathcal{A}\) such that \(\mathcal{A}K\) satisfies [\[eq:23\]](#eq:23){reference-type="eqref" reference="eq:23"}.
Thus, taking subsequence if necessary, there exists a linear transformation \(B \in GL(3)\) such that \(\lim_{m \rightarrow \infty}B_m=B\) in \(GL(3)\), and \(\lim_{m \rightarrow \infty} B_m K_m= BK\) in the Hausdorff distance. Thus \(\lim_{m \rightarrow \infty} \rho_{B_m K_m} = \rho_{BK}\) uniformly in \(S^2 \subset \R^3\). Since \(\Delta_i(B_m K_m)\) and \(\overline{d}(B_m K_m)\), etc. are expressed by the integral of \(\rho_{B_m K_m}\) on closed subsets on \(S^2\), taking as \(m \rightarrow \infty\) in [\[eq:52\]](#eq:52){reference-type="eqref" reference="eq:52"}, we obtain that \[\begin{aligned}
& \abs{\Delta_i(B K)} =\frac{\abs{B K}}{8} \quad (i=1,\dots,4), \\ & \overline{d}(B K) =\overline{e}(B K), \quad \overline{f}(B K) =\overline{g}(B K), \quad \overline{h}(B K) =\overline{i}(B K). \end{aligned}\] Moreover, since each \(B_m K_m\) satisfies [\[eq:23\]](#eq:23){reference-type="eqref" reference="eq:23"}, we obtain \[\begin{aligned}
& \frac{9}{4} |B_m K_m| |(B_m K_m)^\circ| \\ & \geq |Q_1(B_m K_m)|\, |P_1((B_m K_m)^\circ)| \\ &\quad+ |Q_2(B_m K_m)|\, |P_2((B_m K_m)^\circ)|+ |Q_3(B_m K_m)|\, |P_3((B_m K_m)^\circ)| \end{aligned}\] as in Section [3.5](#sec:3.5){reference-type="ref" reference="sec:3.5"}. Since \(\lim_{m \rightarrow \infty} (B_m K_m)^\circ=(BK)^\circ\) in the Hausdorff distance, \[\begin{aligned}
&\frac{9}{4}\mathcal{P}(BK)= \frac{9}{4} |B K| |(B K)^\circ| \\ &\geq |Q_1(B K)|\, |P_1((B K)^\circ)|+ |Q_2(B K)|\, |P_2((B K)^\circ)|+ |Q_3(B K)|\, |P_3((B K)^\circ)| \end{aligned}\] holds. By Theorem [\[prop:12\]](#prop:12){reference-type="ref" reference="prop:12"}, \(|Q_i(B K)|\, |P_i((B K)^\circ)| \geq 8\) \((i=1,2,3)\). Since \(\mathcal{P}(BK)=\mathcal{P}(K)=32/3\), we obtain \(|Q_i(B K)|\, |P_i((B K)^\circ)|=8\) \((i=1,2,3)\). If necessary, we choose a diagonal matrix \(C\) to satisfy that \(\pm (1,0,0), \pm (0,1,0), \pm (0,0,1) \in \partial (CBK)\). Then, \(\mathcal{B}=CB\) satisfies the required conditions. ◻ | {'timestamp': '2017-10-26T02:04:21', 'yymm': '1706', 'arxiv_id': '1706.01749', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01749'} |
# Introduction
Energy injection by supernova explosions can lead to strong outflows from star forming galaxies. These supernovae (SNe) driven galactic outflows could eject a significant fraction of the interstellar medium (ISM) along with metals into their circum-galactic medium (CGM) and the general intergalactic medium (IGM). Feedback due to starburst driven outflows are also invoked in galaxy formation models to get correct shape of the galaxy luminosity functions. It is well demonstrated that a high fraction (40-60%) of high-\(z\) star forming galaxies show signatures of bi-conical outflows in the form of high velocity (i.e. 100-300 km/s) blue shifted absorption in their spectrum. Further, quasar spectra reveal presence of metals in the form of absorption by ions like C iv, O vi etc in the low density intergalactic medium far away from the star forming galaxies [@1996AJ....112..335S; @2002ApJ...578...43C; @2006MNRAS.371L..78R; @2013ApJ...763...37C; @2013MNRAS.435.1198D; @2012MNRAS.421..446M; @2016MNRAS.463.2690D]. Moreover, the CGM has been mapped out to a projected distance of few 100 kpc from the host galaxies using galaxy-quasar pairs. This has revealed the presence of metals in the cold gas at \(10^4\) K (traced by Mg ii and/or Ca ii) as well as in highly ionised warm gas [@2010ApJ...714.1521C; @2011ApJ...740...91P; @2013ApJ...770...41M; @2014ApJ...796..140P] in the CGM. As metals are only produced in stars, they have to be transported into the general IGM or the CGM by large scale galactic outflows. SNe driven galactic outflows have been extensively studied using both hydrodynamical simulations [@2005MNRAS.364..552S; @2006MNRAS.371.1125S; @2008MNRAS.391..110D; @2011MNRAS.415...11D; @2012MNRAS.423.1726S] and semi-analytical models [@1988RvMP...60....1O; @2001ApJ...555...92M; @2002ApJ...574..590S; @2003ApJ...588...18F; @2008MNRAS.385..783S; @2013ApJ...763...17S]. In a previous paper we studied thermally driven outflows from high redshift galaxies and their global effects on the IGM, through a semi-analytic model . Our outflow model was similar to models of stellar wind blown bubbles. Here, the free wind from a galaxy converts its kinetic energy to thermal energy through an inner shock and feeds a hot bubble that expands into the halo medium (or CGM) and subsequently the general IGM (see Fig. [\[wind_profile\]](#wind_profile){reference-type="ref" reference="wind_profile"} below for the entire outflow structure). We showed that outflows can efficiently pollute the CGM and the IGM with metals. However, their detectability in the form of absorption was not clear due to the high temperature and low density of the outflowing gas. Nevertheless, a significant volume of the IGM could be filled with metals due to outflows dominated by low mass galaxies (with dark matter halo mass less than \(10^{10}\) M\(_\odot\)). Strong shocks created by exploding SNe also accelerate highly relativistic charged particles, usually referred to as cosmic rays (CRs). Cosmic rays gyrate along the magnetic field lines and transfer energy and momentum to the thermal gas via Alfvén waves [@1968ApJ...152..987W; @1971ApJ...163..503W; @1969ApJ...156..445K; @1971ApL.....8..189K]. In, we explored the possible influence of cosmic rays generated in the SNe shocks on free wind solutions assuming a constant rate of star formation. The addition of energy and momentum due to cosmic rays helps to drive the free wind at the base of the outflow. It is important especially in low mass galaxies where the gas could lose its thermal energy rapidly due to radiative cooling. Further, such cosmic rays can be reaccelerated at the inner-shock of outflows and add pressure to the hot bubble that drives an outer shock (see Fig. [\[wind_profile\]](#wind_profile){reference-type="ref" reference="wind_profile"}). Also, the cosmic rays in the hot bubble do not cool radiatively and their pressure decreases less steeply with adiabatic expansion compared to thermal gas due to the differences in the adiabatic indices. Therefore, it is important to see how their presence can influence the outflow dynamics. Moreover, there is an increasing interest in exploring the consequences of incorporating CRs in simulations of outflows, which all point to their usefulness in driving outflows ([@2014MNRAS.437.3312S; @2016ApJ...816L..19G; @2016ApJ...824L..30P; @2017ApJ...834..208R; @2017MNRAS.465.4500P];[and also see the review by @2017ARA&A..55...59N section 3.4 which emphasizes the role of CRs]). Thus incorporating the effect of CRs in our semi-analytical outflow models to examine their global consequences for the IGM is also timely. Further, our cosmic ray free-wind solutions showed that the mass loading factor \(\eta_w\), (ratio of mass loss rate due to wind to star formation rate), goes as \(v_c^{-2}\), where \(v_c\) is the circular velocity of the halo. This implies a strong negative feedback on star formation in low mass galaxies. It is then not obvious that such a reduced star formation can indeed drive and sustain strong outflows that go beyond the virial radius in low mass galaxies. Our earlier thermally driven outflow models also did not take into account this negative feedback in star formation. Thus, it is important to revisit outflow models considering not only the effects of cosmic ray driving but also incorporating the negative feedback on star formation implied by such outflows. An important ingredient in outflow models is the star formation rate in a forming galaxy. In a series of papers, we have built up a semi-analytical picture of galaxy formation which reproduces a variety of observations of the high redshift universe. These include UV & Lyman-\(\alpha\) luminosity functions, clustering of Lyman-break galaxies and Lyman-\(\alpha\) emitters, the correlation between star formation rate, stellar mass and halo mass of galaxies. In particular, we use here the models of which incorporated supernova feedback in determining the star formation rate in the high redshift galaxies. There are several additional reasons to study such outflow models. Note that cosmic rays always remain coupled to magnetic fields which are themselves coupled to the thermal gas. Therefore, when winds transport metals into the IGM, they will also carry along magnetic fields and cosmic rays. Thus outflows spread magnetic fields that are being generated/amplified in galaxies to the intergalactic medium. Recent \(\gamma\)-ray observations of distant blazars have even suggested a lower limit to the magnetic fields in void regions at a level of \(10^{-16}\)G if correlated on Mpc scale. have studied possibility of IGM magnetisation by galactic outflows. It is interesting to reconsider this issue with our new improved models of outflows and estimate the strength and volume filling factor of magnetic fields that can be seeded by these outflows into the IGM. Further, if sufficient magnetic fields and cosmic rays are present in the IGM, cosmic rays will still be able to transfer their residual energy to the thermal gas via Alfvén waves. Pockets of over pressured cosmic ray regions can also transfer energy to the IGM via adiabatic heating as they expand and come into pressure equilibrium with their surroundings. If the remaining cosmic ray energy (that is after driving the outflows) is comparable to the thermal energy of the IGM it can potentially influence the thermal history of the IGM. This is interesting as observations have suggested a rise in IGM temperature at \(z\sim 3\) from combining simulations and quasar spectrum . Some authors have attributed this to Helium reionization. However, cosmic rays could also play an important role in heating the IGM, a possibility that we also explore in this work. The paper is organised as follows. In section [2](#sec_outflow){reference-type="ref" reference="sec_outflow"}, we outline our outflow models. Magnetic field evolution is described in section [3](#sec_mag_field){reference-type="ref" reference="sec_mag_field"}. The characteristics of individual outflows is studied in section [4](#sec_wind_profile){reference-type="ref" reference="sec_wind_profile"}. The global impact of outflows is discussed in section [5](#sec_global){reference-type="ref" reference="sec_global"}. The implications of such outflows are addressed in section [6](#sec_impl){reference-type="ref" reference="sec_impl"}. Finally section [7](#sec_dc){reference-type="ref" reference="sec_dc"} presents a discussion of our results and the conclusions. Through out this work we assume cosmological parameters as suggested by Planck results, namely, a flat \(\Lambda\)CDM cosmology with \(\Omega_\Lambda = 0.70\), \(\Omega_b=0.044\), \(n_s=0.96\), \(H_0=71\) km/s/Mpc and \(\sigma_8=0.8\).
# Outflow dynamics {#sec_outflow}
The structural properties of the outflow remain by and large the same as in . However, we improve this model by incorporating the extra pressure due to the non-thermal cosmic ray particles in the outflow dynamics. We assume a spherically symmetric thin shell model of outflows from galaxies. The coherent explosions of SNe in a galaxy produce a bubble of hot gas that expands and expels gas from ISM. The hot bubble expands as the 'free wind' coming from subsequent SNe explosions produces a shock and converts its kinetic energy to thermal energy. The free wind consists of thermal as well as non-thermal particles (CRs) that are produced at the SNe terminal shocks. Such 'free wind' solutions with both thermal and cosmic ray components have been investigated in . The asymptotic velocity of such a free wind is closely related to the circular velocity of the halo. If the halo gas is of sufficient density the free wind produces a shock and converts its kinetic energy to the thermal energy that feeds the hot bubble. We call this as 'inner shock' of the outflow. At this inner shock cosmic rays can also be accelerated through diffusive shock acceleration mechanism with efficiency as high as 50% [@2003ICRC....4.2039K; @2005ApJ...620...44K] in transferring the kinetic energy of the wind to CR particles. The thermal pressure of the gas as well as the cosmic ray pressure in the hot bubble drives an 'outer shock' and sweeps up CGM/IGM material. The swept up material by the outer shock remains confined to a thin shell region and separated from the shocked free wind material by contact discontinuity. In fact, self-similar solution of cosmic ray and thermal gas driven outflow shows that most of the mass of the hot bubble is also expected to be concentrated at the contact discontinuity. See Fig. [\[wind_profile\]](#wind_profile){reference-type="ref" reference="wind_profile"} for the schematic diagram of our outflow models. The evolution of such spherically symmetric outflows is governed by the following equations in the presence of cosmic rays and the hot gas, \[\begin{aligned}
\frac{{\rm d} ^2 R_s}{{\rm d} t^2} &=& \frac{4\pi R_s^2 \left[P_b +P_c-P_0\right]}{m_s(R_s)}-\frac {\dot m_s(R_s)\left[\dot{R}_s-v_0(R_s)\right] }{m_s(R_s)} \nonumber \\ & & \:\:\:\:\:\:\:-\frac {G M(R_s)}{R_s^2}, \label {eqnradius} \end{aligned}\] \[\frac{{\rm d}}{{\rm d} t}\left[m_s(R_s)\right] = \epsilon 4 \pi R_s^2 \rho_B(R_s) \left[\dot{R}_s-v_0(R_s)\right]. \label {eqnmaass}\] Here, the position of the outer shock is \(R_s\), \(m_s\) is the gas mass of the thin shell and \(P_b\) and \(P_c\) are the thermal and cosmic ray pressures in the hot bubble. Further, \(\rho_B\) and \(P_0\) are respectively the baryon density and the pressure of the outside medium assuming virial temperature for halo gas or \(10^4\) K for the IGM; \(v_0\) is the ambient velocity field taken from. The first term in the right hand side of Eq. [\[eqnradius\]](#eqnradius){reference-type="ref" reference="eqnradius"} represents the net pressure i.e. thermal plus cosmic ray pressure minus the outside pressure, that drives the outer shock. The second term is due to the mass loading of the shell from the outside medium causing the deceleration of the shock and we assume a fraction \(\epsilon\) of all swept up mass remains in the thin shell (i.e. Eq. [\[eqnmaass\]](#eqnmaass){reference-type="ref" reference="eqnmaass"}) and rest is mixed with the hot bubble due to evaporation/fragmentation (the mass in the hot bubble is negligible compare to the swept up mass). The 3rd term in Eq. [\[eqnradius\]](#eqnradius){reference-type="ref" reference="eqnradius"} is the gravitational attraction due to the dark matter [a NFW profile is assumed with concentration parameter \(c=4.8\), @1997ApJ...490..493N; @2001ApJ...555...92M] as well as baryonic matter inside the radius \(R_s\). For the initial baryonic density profile inside the halo we assume a beta model and assume that a fraction, \(f_h=0.1\) of total the total gas remains in the halo. The thermal pressure \(P_b\) can be obtained from the thermal energy of the bubble, \(E_b\), by \[P_b = \frac{E_b}{ 2\pi (R_s^3-R_1^3)}, \label{pressure_energy}\] where we assume that the adiabatic index of the thermal gas \(\gamma = 5/3\). Here, \(R_1\) is the position of the inner shock of the outflow. Similarly, the non-thermal cosmic ray pressure \(P_c\) and the cosmic ray energy in the bubble, \(E_c\), are related by, \[P_c = \frac{U_c}{3}=\frac{E_c}{ 4\pi (R_s^3-R_1^3)}, \label{pressure_energy_cr}\] with the assumption that the adiabatic index of the CR component \(\gamma_{CR} = 4/3\) (considering them as highly relativistic particles). Here, \(U_c\) is the energy density of the cosmic ray particles. The evolution of the thermal energy of the bubble is given by, \[\frac {{\rm d} E_b}{{\rm d} t} = L(t)-\Lambda (t,T,Z) -4\pi \left[R_s^2 \dot{R}_s-R_1^2 \dot{R}_1\right] P_b. \label {eqnenergy}\] The 'free wind' carries some fraction of total SNe energy in the host galaxy which is fed into the bubble as thermal energy, \(L(t)\), at the inner shock. The bubble can lose its thermal energy due to \(P dV\) work as it expands and radiative cooling. For cooling rate i.e. \(\Lambda (t,T,Z)\), we consider Compton drag against the CMBR, bremsstrahlung and recombination line cooling that depends on temperature \(T\) and metallicity, \(Z\) of the gas. The gas density that determines the temperature of the hot bubble is taken to be the average density of the bubble. Note that the bubble is filled by the gas coming from the free wind and a fraction of the swept up mass. The free wind mass is higher for low mass galaxies owing to higher mass loading factor, \(\eta_w\), that leads to higher radiative cooling of the bubble. Further, the metallicity of the bubble gas is self-consistently determined as described in the Appendix A of . Similarly, the evolution of the cosmic ray energy in the bubble is governed by \[\frac {{\rm d} E_c}{{\rm d} t} = L_c(t)-4\pi \left[R_s^2 \dot{R}_s-R_1^2 \dot{R}_1\right] P_c. \label {eqnenergycr}\] Here, \(L_c\) is the total energy of the cosmic ray component per unit time injected into the bubble, at the inner shock. Note we consider, at the inner shock, the cosmic rays are being accelerated through diffusive shock acceleration with efficiency as high as 50% in converting the kinetic energy of the shock front into cosmic rays. Thus, if initially we have 10% of total supernova energy available in the free wind, at the inner shock half of it will be available as thermal energy and rest half will be converted into cosmic ray energy. It is interesting to note that cosmic rays lose energy only due to adiabatic expansion that is solely determined from the outflow dynamics. But the thermal energy is lost by additional cooling processes and can be significant fraction of its initial energy. Thus, in case of only thermally driven wind one could encounter a situation where the gas has lost most of its thermal energy, the inner shock catches up with the outer shock and only the momentum of the free wind drives the outflow. This is known as 'momentum driven outflow' [see @2008MNRAS.385..783S for evolution of such outflows]. However, in the presence of non-zero cosmic ray energy, we will find that there is always a finite separation between the inner and outer shocks and hence the outflow never transits to the momentum driven case. Further, owing to a softer equation of state, cosmic rays lose energy slower than the thermal gas during adiabatic expansion. Hence even when radiative cooling of the bubble gas is efficient, cosmic ray pressure is important and moreover it dominates the outflow dynamics at the later stages of its evolution. Most analytical models of galactic outflows do not consider the dynamics of the inner shock at \(R_1\) . We follow its dynamics using simply the jump condition across the inner shock boundary in a two fluid model assuming a strong shock. The evolution equation for \(R_1\) is given by \[P_b +P_c = \frac{2}{\gamma_s+1} \frac{\dot{M}_w(t_e) }{4 \pi R_1^2 v_w}\left[ v_w-\dot{R}_1\right]^2, \label{R1eqn}\] with \[\frac{2}{\gamma_s+1}=\frac{3\left(1+2\frac{P_{c}}{P_b}\right)}{4+7\frac{P_{c}}{P_b}}. \label{gammas_eqn}\] Here, \(v_w\) is the asymptotic speed of the free wind material before it encounters the inner shock and related to the SNe energy by \(L_0(t)=\dot{M}_w(t) v_w^2/2\), with the mass outflow rate \(\dot{M}_w(t)\) is obtained from Eq. [\[eqn_eta\]](#eqn_eta){reference-type="ref" reference="eqn_eta"}. The mechanical luminosity, \(L(t)\), and the cosmic ray luminosity at the inner shock, \(L_c(t)\), fed into the wind bubble are related to the SNe explosion energy and hence the star formation rate of the host galaxy. Our model for calculating this star formation rate is described in Appendix [\[sec_sfr\]](#sec_sfr){reference-type="ref" reference="sec_sfr"}. It takes into account the negative feedback due to mass loss in the free wind following. However, we make some modifications that are needed in order to get a more accurate time resolved star formation rate in the low mass galaxies. The resulting evolution of the SFR for a few sample halo masses and collapse redshifts incorporating SNe feedback, are shown in Fig. [\[fig_sfr\]](#fig_sfr){reference-type="ref" reference="fig_sfr"}, where we can see a significant suppression of the SFR for low mass galaxies. This model of star formation is similar in philosophy to the 'bathtub model' of and, applied however just after the formation of a new dark matter halo, either formed by mergers or from a collapsing density peak. Thus only the first burst of star formation lasting several halo dynamical time scale is considered. It does not consider the slow and continuous star formation mode of a galaxy that arises at a later stage of its evolution due to slow accretion process. Note that such a slow star formation may not result in an outflow that escapes the dark matter potential. We will explore such effects in future, but concentrate for now on the major epoch of star formation. The luminosities \(L(t)\) and \(L_c(t)\) are related to the star formation rate \(dM_*/dt\) by, \[L(t) = 10^{51}~{\rm ergs}\times\epsilon_{\rm w}~\nu~\frac{dM_*}{dt}\] and \[L_c(t) = 10^{51}~{\rm ergs}\times\epsilon_{\rm cr}~\nu~\frac{dM_*}{dt}.\] Here, we assume that \(\nu\) number of SNe are formed per unit mass of star formation and each SNe produces \(10^{51}\) ergs of energy on an average. A fraction of that energy, \(\epsilon_w\) is channeled into the hot bubble as thermal energy and a fraction, \(\epsilon_{\rm cr}\), is transformed in to the cosmic ray energy. In our model where the SNe first drives a free wind, which then flows into the hot bubble through a strong inner shock, it is this strong shock which reconverts the kinetic energy of the free wind back into both thermal energy and cosmic rays. Thus, the total energy efficiency in converting SNe energy to energy available for driving the outflow is \(\epsilon_w + \epsilon_{\rm cr}\). For most of our calculations we assume \(\epsilon_w=0.05\) and \(\epsilon_{\rm cr}=0.05\) making total energy efficiency of 0.10. This efficiency is 2-3 times lower than the efficiency that has been found in numerical simulations (see where they obtained 20-30% efficiency) and hence a conservative value. Further, we assume a Salpeter initial mass function for the distribution of formed stars in the mass range \(1-100\) M\(_\odot\) that results in one SNe per 50 M\(_\odot\) of star formation. The above equations are simultaneously solved numerically starting from a set of initial conditions as described in . We follow the outflow dynamics till the outflow peculiar velocity decreases to the local sound speed in the IGM. At this point we assume that the shock would disappear and the outflowing material would mix with the IGM and expand with the Hubble flow.
# Magnetic field dynamics {#sec_mag_field}
It is important to explore the magnetisation of the IGM by outflows. We have already mentioned that the magnetic fields play an important role in its interaction with cosmic rays and hence probably in the thermal history of the universe. Galactic outflows are potential candidates to magnetise the IGM. Here, we wish to get a conservative estimate of the strength of the IGM magnetic fields resulting from our outflow models. For simplicity, we treat evolution of magnetic fields separately and do not consider any dynamical effects of magnetic fields on the evolution of the outflow. Note that, adding the dynamical effect of magnetic fields can further aid in driving outflows. However, we have adopted a more conservative approach in present work as the magnetic pressure is always significantly subdominant to the CR pressure. We assume that micro-Gauss level magnetic fields are present in the ISM of the high redshift protogalaxies when outflows begin. These fields themselves would be generated by turbulent dynamo processes in the galaxy. The ionised plasma in the wind would be coupled with this magnetic field and transport the magnetic flux out of the galaxy with the outflowing material. In order to calculate the strength of magnetic fields that can be deposited in the IGM via this process we follow where they considered magnetising the IGM by thermally driven winds. There are two possible scenarios in the co-evolution of magnetic fields and outflows. In the first case that we refer as 'conservative' model, the magnetic fields from galaxies are injected into the outflowing material and just get diluted as they expand with the outflow, obeying the magnetic flux freezing condition. The evolution of magnetic fields in such models is governed by following equation, \[\frac{dE_B}{dt}=\dot{E}_{B_{in}}-\frac{1}{3}\frac{\dot{V}_w}{V_w}E_B.\] Here, \(E_B\) is the total magnetic energy in the wind bubble; \(V_w\) is the volume of the wind bubble. Further, \(\dot{E}_{B_{in}}\) is the magnetic energy injection rate in the bubble from the galaxy and we take \[\dot{E}_{B_{in}}=\epsilon_{B_{in}}\frac{\dot{M}_w}{\bar{\rho}_{in}} =\epsilon_{B_{in}} 4\pi R_1^2 v_w\] with the density of the free wind material is given by \[\bar{\rho}_{in}=\frac{\dot{M}_w}{4\pi R_1^2 v_w}\] and the magnetic energy density that is injected in the wind bubble is \[\epsilon_{B_{in}}=\frac{B^2}{8\pi} \left(\frac{\bar{\rho}_{in}}{\bar{\rho}_{\rm ISM}}\right)^{4/3}. \label{eqn_B}\] \(B\) is the magnetic field inside the star forming galaxy. The value of \(B\) ranges from \(\sim 10-20\mu\)G in nearby spirals to values of \(B\sim 50-100\mu\)G in nearby starburst galaxies and in barred galaxies. There is also tentative evidence from high-\(z\) Mg ii system that \(z\sim 1\) galaxies are already magnetized to current levels. Thus it seems reasonable to adopt \(B\sim 10-20\mu\)G for the galaxies at high redshift which are driving outflows. Further, \(\rho_{\rm ISM}\) is the density of the ISM gas taken as 1000 times the average baryonic density of the halo. In the above 'conservative' case we ignore the possible amplification of magnetic fields inside the outflow and thus we get the minimum possible value of magnetic fields that can be inputted into the IGM via outflows. In a more optimistic model B we take into account possible amplification of magnetic fields in the wind plasma due to shear flow and turbulence of the wind material. The characteristic time scale for this process is roughly given by, \[\tau_{\rm{ eff}}^{-1} = f_s\frac{\dot{R}_s-H(t)R_s}{R_s} = f_s\left(\frac{\dot{R}_s}{R_s}-H(t)\right) \label{taueff}\] with \(H(t)\) is the Hubble parameter. Here, \(\dot{R}_s-H(t)R_s\) is the peculiar velocity of the outflow with respect to Hubble flow and dividing it by \(R_s\) gives a measure of the velocity shearing rate in the outflowing material. Such a velocity could reflect itself in turbulence, which amplifies the field, due to a small-scale turbulent dynamo, on the turbulent eddy turn over timescales. In such a dynamo the smallest eddy which has a magnetic Reynolds number above a critical value decides the rate of initial amplification. The fudge factor \(f_s > 1\) gives how much faster such an amplification occurs with respect to the shear time-scale. It will depend on the poorly known properties of the turbulence in the hot bubble and could be much greater than unity, if the smallest super critical eddy is much smaller than the bubble size. We assume \(f_s \sim 1\) as a conservative estimate below. Including this amplification, the change in magnetic energy can be calculated from \[\frac{dE_B}{dt} = \dot{E}_{B_{in}}+ \left(\frac{1}{\tau_{\rm{eff}}}-\frac{1}{3}\frac{\dot{V}_w}{V_w}\right)E_B\] which for \(R_s \gg R_1\), can be simplified to \[\frac{dE_B}{dt} = \dot{E}_{B_{in}}-H(t)E_B.\] This would provide us a more optimistic value of the magnetic fields with which IGM can be seeded via outflows.
# Structural characteristics of outflows {#sec_wind_profile}
Having set up the machinery for the evolution of outflows from star forming galaxies, driven by the thermal and non-thermal pressures and the spreading of magnetic fields and cosmic rays via such outflows, we present here some important characteristics of individual outflows resulting from our numerical calculations. We first focus on two representative masses of host galaxies in order to show the general trends in our models.
## Outflow from dwarf galaxies
We begin with the study of outflows from a host galaxy of total mass \(M=10^8\) M\(_\odot\) that has collapsed at \(z_c=10\). The star formation history of such galaxies is shown by the solid red curve in Fig [\[fig_sfr\]](#fig_sfr){reference-type="ref" reference="fig_sfr"}. As we can see from this figure, in such small mass galaxies, due to strong negative SNe feedback or a large \(\eta_w\), star formation completely stops after \(10^7\) yrs. Hence, it is essential to investigate whether such strongly suppressed and brief duration of star formation can at all produce enough SNe energy to drive a galactic scale outflow from the galaxy. The important characteristics of the outflow originating from this particular galaxy are shown in Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"}. Indeed we see in Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"} that a galactic scale outflow has resulted from such galaxy. It is clear from panel (a) of Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"}, where we plot the location of inner and outer shocks as a function of time, that the outflow crosses the virial radius of the halo at \(\sim 3\times 10^7\) yrs, goes far beyond the virial radius (shown by horizontal line) before it mixes with the IGM. The final physical radius of the outflow is 20 kpc before it is frozen into the Hubble flow after \(\sim 3\times 10^8\) yrs from the collapse of the halo (as indicated by the vertical lines in panels (a) and (b) of Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"}). Note that the halo has a virial radius of 1.4 kpc. Therefore, the outflow originating from the host galaxy is able to pollute IGM with metals, magnetic fields and CRs upto \(\sim 15\) times of its virial radius. The outflow achieves a maximum speed of 110 km/s just after crossing the virial radius of the halo (panel (b) and (d) of Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"}). The sudden increase in the velocity after crossing the virial radius is due to rapid change in the assumed gas density profile outside the halo. It is interesting to note that even though star formation ceased after few times \(10^7\) yrs the inner shock still persists upto a time \(\sim 5 \times 10^7\) yrs (panels (a) and (c)). This is due to the fact that it takes a finite time for the effects of energy release by SNe to be felt at the inner shock. The inner shock starts from the centre of the galaxy (our initial condition), rapidly goes closer to the outer shock reaching to a maximum of \(0.65 R_s\). Thus, it never comes very close to the thin shell and the outflow never transits to a momentum driven stage. It can be seen from panel (f) of Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"} that the temperature of the bubble is always less than \(10^4\) K. So in the absence of CRs, the thermal pressure alone would not have sustained the bubble region leading to a direct momentum impact of the free wind to the thin shell. However, in presence of cosmic ray pressure this never happens. At a very initial stage of the flow the thermal pressure and cosmic ray pressure were comparable to each other. But soon the gas loses its thermal energy by radiative cooling due to higher density of the wind material. At this point, the comic ray pressure becomes an order of magnitude higher than the thermal pressure even if initially they were of same order (see panel (e) of Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"}). This extra cosmic ray pressure completely governs the further evolution of the outflow keeping it always like an 'energy driven flow'. We now turn to the spreading of metals and magnetic fields by the outflow. Panel (g) of Fig. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"} shows the magnetic field strength in the outflowing gas, as a function of the outflow radius. Here, we have used the 'conservative' model for the magnetic fields evolution, and assumed \(B\sim 10 \mu\)G in the galaxy. By the time outflow reaches the virial radius, the average strength of the magnetic fields is \(\sim 0.02~\mu\)G. It reduces farther reaching to a value of \(0.4~\)nano Gauss when the outflow mixes with the IGM. This field strength could be even larger if we take larger \(B\) for the galactic field as would be appropriate for a star bursting galaxy. Note that the 'optimistic' model would predict an order of magnitude higher magnetic field strength. The implications of such fields will be discussed in Sec 6.2 and 6.3. The amount of metals carried by the wind material from the ISM to the bubble increases with time and hence it increases the metallicity of the bubble gas. We see from panel (h) that by this outflow we can pollute IGM with a metallicity of \(Z\sim 0.001\) Z\(_\odot\). In a nutshell, the star formation and resulting SNe explosions in a \(10^8\) M\(_\odot\) halo can produce a galactic scale outflow of 'cold' gas, mostly driven by the cosmic ray pressure and can pollute a spherical region of radius 20 kpc (i.e. \(\sim 15\) times the virial radius) with metals (metallicity of the polluted region is typically \(\sim 0.001\) Z\(_\odot\)), cosmic rays and magnetic fields with of order a nano Gauss strength when it mixes with the IGM.
## Outflow characteristic of a galactic scale halo
We now consider the outflow from a more massive halo with mass \(M=10^{11}\) M\(_\odot\) and collapse redshift of \(z_c=4\). The outflow properties are shown in Fig. [\[fig_profile11\]](#fig_profile11){reference-type="ref" reference="fig_profile11"}. We choose such an example as we will show later that these halos contribute most to the global volume filling of the wind material today. The basic picture remains the same as its low mass counterpart. However, there are certain differences. Owing to its higher star formation rate, the outflow from the halo extends to more than a Mpc from the centre of the galaxy before freezing into the Hubble flow (panel (a)). It achieves a maximum velocity of 340 km/s while crossing the virial radius of the halo (panel (b) and (d)). The outflow remains 'energy driven' through out its lifetime; the inner shock never catches up the outer shock as can be seen in panels (a) and (c) of Fig. [\[fig_profile11\]](#fig_profile11){reference-type="ref" reference="fig_profile11"}. The most striking difference between outflows from high and low mass galaxies is in the temperature evolution. In the low mass galaxy the outflowing material is cold. But in the high mass galaxy the density of hot bubble is low due to smaller \(\eta_w\) making the radiative cooling less efficient. Thus as can be seen in panel (f) of Fig. [\[fig_profile11\]](#fig_profile11){reference-type="ref" reference="fig_profile11"}, the temperature of the hot gas is more than \(10^6\) K when the outflow starts. It decreases afterwards, but never becomes less than \(10^4\) K when the outflow is in active stage. Further, the thermal and cosmic ray pressures remains comparable to each other through out the life time of the outflow (panel (e)). It is interesting to understand the evolution of both the pressures. Initially, the thermal pressure is higher than the cosmic ray pressure. However, after some time the cosmic ray pressure becomes higher. Such a feature arises due to different adiabatic indices for thermal gas and relativistic cosmic ray component causing pressure of the non-thermal component to decrease more slowly under adiabatic expansion (see Eqs. [\[pressure_energy\]](#pressure_energy){reference-type="ref" reference="pressure_energy"} and [\[pressure_energy_cr\]](#pressure_energy_cr){reference-type="ref" reference="pressure_energy_cr"}). As can be seen from panel (h) of Figs. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"} & [\[fig_profile11\]](#fig_profile11){reference-type="ref" reference="fig_profile11"}, \(10^{11}\) M\(_\odot\) halo spreads the metals over large distances and the resulting gas phase metallicity in the polluted region can be two orders of magnitude higher than that of a \(10^{8}\) M\(_\odot\) galaxy. The large galaxy produce more metals through SNe and thus eject more via outflow. On the other hand, they both magnetise the IGM with similar magnetic field strength but over different volumes. Thus outflows from massive galaxies pollute a fairly large volume of IGM with hot gas having higher metallicity. But they contribute in a similar fashion in magnetising the IGM. This may change if we consider a halo mass (or SFR) dependence in \(B\) of Eq. [\[eqn_B\]](#eqn_B){reference-type="ref" reference="eqn_B"}. Further, note that in panel (g) of Figs. [\[fig_profile\]](#fig_profile){reference-type="ref" reference="fig_profile"} and [\[fig_profile11\]](#fig_profile11){reference-type="ref" reference="fig_profile11"}, the evolution of the magnetic fields has been obtained for our 'conservative' models. The more 'optimistic' model B produces an order of magnitude higher magnetic field in the IGM. We will discuss its implication on the thermal history of IGM in section 6.2.
## Comparison with Previous models
In this section we compare both cosmic ray and thermally driven outflows with models of that considered only the thermally driven outflows and a fixed mass loading factor \(\eta_w=0.3\). As already mentioned, there are several improvements in present models compared to the model presented in . Firstly, the prescription for the star formation now takes into account the negative feedback arising from the SNe driven winds. Secondly, we also take into account the increased mass loading factor \(\eta_w = (v_c/v_c^*)^{-2}\) in small galaxies in the outflow dynamics as adopted in our star formation model. Finally, the winds are now driven by both thermal and cosmic ray pressures. As described before, the first two effects have opposite influence on the wind dynamics compared to the last one. In this section we compare different models in more detail. In Fig. [\[fig_comp\]](#fig_comp){reference-type="ref" reference="fig_comp"}, we show the properties of thermally driven outflows with models discussed in by blue long dashed curves. The corresponding curves for models that includes SNe feedback, the increased mass loading but no effect of cosmic rays, are given in short dashed dark-green lines. In both the cases, we assume \(\epsilon_w=0.1\) and \(\epsilon_{\rm cr}=0\). The effect of the SNe feedback, increased mass loading, but now including the cosmic rays in the outflow dynamics keeping the total efficiency the same (i.e. \(\epsilon_w + \epsilon_{\rm cr} =0.1\)), is shown as red solid lines. The bottom panels consider outflow from a \(10^8\) M\(_\odot\) galaxy while the top panels are for \(10^{11}\) M\(_\odot\) halo. It is clear from the figure that the outflow escapes the low mass galaxy for the constant \(\eta_w\) model of . However, including supernova feedback in star formation and increased \(\eta_w\) prevent the outflow from taking off in such galaxies. This is reflected by the fact that the green short dashed curves are hardly noticeable in the first two columns of bottom panels of Fig. [\[fig_comp\]](#fig_comp){reference-type="ref" reference="fig_comp"}. This is because of (i) the reduction in the star formation (especially due to the larger \(\eta_w\) in dwarf galaxies) and (ii) increase in the mass loading that leads to higher density and hence rapid cooling of the hot bubble gas. It is also evident from the figure that one can have outflows in such halos when we include the cosmic rays in the outflow dynamics (solid red curves) keeping the total energy efficiency the same. Thus we conclude that the presence of CRs enables outflows even in small mass galaxies with a large mass loading. Further, the outflowing gas will have low temperature due to efficient radiative cooling (temperature is \(< 10^4\) K as can be seen in the left bottom panel of Fig. [\[fig_comp\]](#fig_comp){reference-type="ref" reference="fig_comp"}). For large galaxies with \(M=10^{11}\) M\(_\odot\), outflows do escape the dark matter potential in all three cases as described above. However, the cosmic ray driven outflows with same energy efficiency reaches much larger distance (see top panels of Fig. [\[fig_comp\]](#fig_comp){reference-type="ref" reference="fig_comp"}). This is because the cosmic ray pressure drops slower with radius than the thermal pressure due to a different adiabatic index. Thus cosmic ray pressure is able to push the outflowing gas further away compared to the one driven by thermally driven outflows. This can be seen from the velocity profile as plotted in the top middle panel of Fig. [\[fig_comp\]](#fig_comp){reference-type="ref" reference="fig_comp"}. Further, the temperature of the bubble gas is also lower by an order of magnitude in case of cosmic ray driven outflows making it more plausible for detection of metals.
# Global influences of outflows {#sec_global}
In previous section we have investigated several key characteristics of outflows originating from the high redshift star forming galaxies. Here we show the global impact of outflows on the IGM at different redshifts. The most interesting physical quantity is the total volume of the universe/IGM that is affected/polluted by these outflows. In this respect we define volume filling factor as \(F=1-\exp [-Q(z)]\) with porosity \(Q\) is defined as \[\begin{aligned}
Q(z) = \int \limits_{M_{\rm low}}^{\infty} {\rm d} M \int \limits_z ^{\infty} {\rm d} z' \frac{{\rm d}^2 N(M,z,z')}{{\rm d} z'{\rm d} M} \frac{4}{3}\pi \left[R_S(1+z)\right]^3. \label{eqn_Q} \end{aligned}\] Here, \({{\rm d}^2 N(M,z,z')}/{{\rm d} z'{\rm d} M}\) is the comoving number density of dark matter halos having mass between \(M\) and \(M + {\rm d} M\) that are formed in redshifts range \(z_c\) and \(z_c + {\rm d} z_c\) and surviving till redshift \(z\) without being merged with other halo. We use the modified Press-Schechter (PS) formalism of to calculate this number density. In particular, it is obtained from \[\begin{aligned}
\frac{{\rm d}^2 N(M,z,z')}{{\rm d} z'{\rm d} M}~ {\rm d} M~ {\rm d} z' &=& N_M(z') \left(\frac{\delta_c}{D(z') \sigma(M)}\right)^2 \frac{\dot{D}(z')}{D(z')}\;\nonumber \\ & \times & \frac{D(z')}{D(z)} \frac{{\rm d} z'}{H(z') (1 + z')}~ {\rm d} M. \label{eqnmPS} \end{aligned}\] Here, \(N_M(z') {\rm d} M\) is the Press-Schechter mass function, \(\delta_c=1.686\) is the critical over density for collapse, \(H(z)\) is the Hubble parameter, \(D(z)\) the growth factor for linear perturbations and \(\sigma(M)\) the rms mass fluctuation at a mass scale \(M\). Note that Eq. [\[eqnmPS\]](#eqnmPS){reference-type="ref" reference="eqnmPS"} does not take into account the survival of the hot bubble when two halos merge to form a bigger halo and thus using it in Eq. [\[eqn_Q\]](#eqn_Q){reference-type="ref" reference="eqn_Q"} would provide a lower limit on \(Q\) or \(F\). On the other hand if we use simple derivative of any mass function as the formation rate of halos and use it as \({{\rm d}^2 N(M,z')}/{{\rm d} z'{\rm d} M}\), we would get an upper limit on the \(F\) or \(Q\) as we do not consider the merging of outflows. To explore this we use derivative of Sheth-Tormen mass function to calculate the formation rate of halos and hence the volume filling factor and associated properties of outflows. The lower mass limit (\(M_{\rm low}\)) in Eq. ([\[eqn_Q\]](#eqn_Q){reference-type="ref" reference="eqn_Q"}) is determined from the physical conditions required to host star formation. Before reionisation, gas inside a dark matter halo can cool via atomic hydrogen if virial temperature (\(T_{\rm vir}\)) of the halo is greater than \(10^4\) K. In the presence of molecular hydrogen, a halo with virial temperature as low as 300 K can cool and host star formation. In what follows we consider two models: (i) atomic cooled models where we assume \(M_{\rm low}\) corresponds to \(T_{\rm vir}=10^4\) K and (ii) molecular cooled models where we take \(M_{\rm low}\) corresponds to \(T_{\rm vir}=300\) K. Further, in the ionised regions of the universe due to radiative feedback, a halo can host star formation only if its circular velocity is more than about 35 km/s due to increase of Jeans mass. Thus in order to know whether a halo can host star formation or not we need to follow the reionisation history of the universe self-consistently. We follow in order to calculate the ionisation state of the universe. At this point we wish to note that all our self-consistent reionisation models are compatible with available observations, namely reionization redshift \(z_{re}\gtrsim 6\) and the resulting electron scattering optical depth is within \(\tau_e=0.058\pm 0.012\). Further for any physical quantity related to outflows, we calculate porosity weighted average as \[\begin{aligned}
\langle X\rangle &=& Q^{-1} \int \limits_{M_{\rm low}}^{\infty} {\rm d} M \int \limits_z ^{\infty} {\rm d} z'~ \frac{{\rm d}^2 N(M,z,z')}{{\rm d} z'~{\rm d} M}\times \nonumber \\ & & ~~~~~~~~~~~~~~~~~~~~~~~~~~~\frac{4}{3}\pi \left[R_S(1+z) \right]^3 X, \label{eqnavg} \end{aligned}\] \(X\) being the physical quantity. If one does not include \(Q^{-1}\) factor it would correspond to the mean value of \(X\) in the IGM.
## Atomic cooling models
We start by showing results for our 'atomic cooled' models with Press-Schechter mass function. In Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"} we show various physical quantities related to outflows as a function of redshift by solid curves. In panel (a) we show the most relevant quantity, the volume filling factor \(F\) of the bubbles. Initially at \(z=10\) the filling factor is small \(\sim 0.01\), gradually increases to 0.26 at \(z=3\) and finally reaches to 0.6 by \(z=0\). Thus a minimum 60% of the universe can be filled with outflows from atomic cooled halos by today. The average halo mass contributing to the volume filling is plotted in panel (b). Above \(z=6\) the average mass of haloes contributing most in volume filling is \(\sim 10^8\) M\(_\odot\). During this period the average bubble radius is 100 kpc (panel (c)). At later times both the average mass and bubble size increase slowly due to two reasons. Firstly, the number of higher mass galaxies are increasing in a hierarchical structure formation scenario. Secondly, the existing outflows are getting more time to grow in size. At \(z=0\), massive galaxies with average mass \(\sim 10^{11}\) M\(_\odot\) having outflows of Mpc scale are responsible for polluting the IGM. The porosity weighted peculiar velocity at different redshift is plotted in panel (d) of Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"}. Initially, most of outflows are in the early active stage having higher individual outflow velocity; thus the average peculiar velocity is also as high as 50 km/s. However, the outflows are coming from small mass galaxies. Hence they have a lower average temperature of \(10^3\) K (see panel (e)). As the time passes, these outflows grow older and their speed reduces. But as new massive galaxies started collapsing, the peculiar velocity raises to 55 km/s at \(z=2\). The average temperature of the bubble also shows a peak of \(10^5\) K at the same time. The average density of the bubble is shown in panel (f) by the solid curve and most of the time it stays bellow the average IGM density (shown by the dashed line) at that redshift. This is expected as the hot bubble sweeps up the IGM into a shell creating a low density, high temperature regions. In passing we note that several porosity weighted average physical quantities do not show monotonic behaviour with redshift such as porosity weighted peculiar velocity and radius of outflows. A number of counter veiling effects are responsible for such non-monotonic behaviour of the porosity weighted peculiar velocity (\(v_p\)). Although the characteristic mass of collapsing halos increases monotonically with decreasing redshift, the corresponding stellar to halo mass ratio is not monotonic with mass because of the reionization, supernovae and AGN feedbacks. Also as outflows expand, their velocity decreases, which is reflected by corresponding decrease in porosity weighted \(v_p\) for \(z > 6\). Then after reionization, small mass halos lose their importance, more and more massive galaxies with younger outflows start to dominate and increase the average peculiar velocity till \(z \sim 2\). After this star formation from very massive galaxies is being suppressed due to AGN feedback, while outflows from normal galaxies are becoming older, and the weighted \(v_p\) decreases again. One of the reasons for studying outflows is to understand how efficiently they can pollute IGM with heavier elements, magnetic fields and CRs. The metallicity and magnetic field evolution of the IGM are shown in the panels (g) and (h) of Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"}, respectively. As expected the IGM metallicity level increases with time as more and more SNe explode and resulting metals are put into the IGM via outflows. Finally, the metallicity of the IGM that can be reached by today is \(\sim 10^{-2}\) Z\(_\odot\) (panel (g)). Note that these metals are distributed in 60% volume of the universe. In the same volume, the average magnetic field strength that can be seeded by outflows is \(0.4-4\) nano Gauss at \(z=0-3\), for an assumed \(B\sim 20\mu\)G in galaxies, for our conservative model. Models with optimistic magnetic fields evolution predict, \(\langle B\rangle \sim 2-20\) nano Gauss, or an order of magnitude higher magnetic fields seeded by this mechanism, as can be seen from the short dashed line of panel (h) in Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"}. Finally, it is interesting to investigate the magnitude of the porosity weighted cosmic ray energy density that is still remaining in the plasma. We show this in panel (i) of Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"}. Before, \(z= 6\) the remaining CR energy density is as greater than \(10^{-3}\) eV cm\(^{-3}\). However, only 10% of the universe is filled with these cosmic rays. Even at low redshift the energy density of cosmic rays does not reduce much. At \(z=3~(1)\) the excess energy density of in the cosmic rays is \(5\times 10^{-4}\) eV cm\(^{-3}\) (\(10^{-4}\) eV cm\(^{-3}\)) in 26% (55%) volume of the universe. We will discuss this further in the following section and compare this energy density with the thermal energy density of the IGM. We note that there is not much qualitative change in the globally averaged physical properties associated between our current model, where outflows are driven by both thermal and cosmic ray pressure, compared the models of , where we considered outflows to be only thermally driven. The crucial difference however, is that the current models have taken into account also the negative feedback on star formation due to SNe driven mass loss. One property which does change is the overall volume filling factor, which is smaller in the current models, compared to , because of the overall reduction in star formation due to SNe feedback. Further, the average temperature of the outflows has reduced dramatically, due to enhanced mass loading in the hot bubble, which will help in the detectability of the heavier elements in the quasar absorption spectrum.
## Molecular cooled models
We now turn to the 'molecular cooled' models. As discussed in outflows from atomic cooled halos can potentially disturb Lyman-\(\alpha\) forest due to their large peculiar velocity and higher temperature and we showed that molecular cooled halos provide a respite from this by filling the IGM at very early epoch with smaller outflows. Note that, in such small mass galaxies the mass loading factor \(\eta_w\) is so high that the star formation ceases with the onset of supernova explosion, after \(t_{\rm SNe} \sim 3\times 10^7\) yr. Within that period we find that in such galaxies the fraction of baryonic mass turned into stars (i.e. \(M_*/M_b\)) is of order 0.05 at \(z=10\). Thus only about 5% of the baryons is converted to stars in such molecular cooled galaxies at high-\(z\). Nevertheless, this does have a significant effect on the volume filling at high redshift. In Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"} we show the global properties of outflows for such models with long dashed lines. As expected, in these models outflows volume filled quite early; at \(z=8\), more than 75% of the IGM is filled by the wind material and hence with metals, cosmic rays and magnetic fields. The contribution comes mostly from galaxies of masses \(10^7\) M\(_\odot\) or less with average bubble size less than 30 kpc (see panels (b) and (c)). Upto \(z=3\) the average peculiar velocity of the outflowing material is less than 20 km/s with average outflowing gas temperature less than \(10^4\) K. Thus, cold gas from the small mass molecular cooled halos dominates the volume filling of the IGM and is expected not to disturb the Lyman-\(\alpha\) forest. Note our results for molecular cooled halos will depend very much on the assumed value of \(M_*/M_b\). This quantity is poorly constrained for such halos and some models suggest that it could be much less than what we assume here. Understanding star formation efficiency in such low mass galaxies at early epoch is essential to accurately model their effect on global properties. The other physical quantities such as gas density of bubbles, amount of metals and magnetic fields put in the IGM remain similar (panels (f), (g) and (h) of Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"} respectively) in both molecular cooled and atomic cooled models. Note that in Fig. [\[fig_profileg\]](#fig_profileg){reference-type="ref" reference="fig_profileg"} we have shown the possible magnetisation of the IGM with our optimistic models as well and that resulted one order of magnitude higher value compared to conservative model. The excess porosity weighted cosmic ray energy density remains one order of magnitude smaller in case of molecular cooled models for \(z\gtrsim 6\). After that the difference between atomic and molecular cooled models decreases and finally by \(z \lesssim 2\) they become equal.
## Effect of mass function and formation rates
Here, we show effects of different halo mass functions on global properties of outflows. In Fig. [\[fig_PS_STD\]](#fig_PS_STD){reference-type="ref" reference="fig_PS_STD"} we compare the porosity and the volume filling factor as obtained from using Sheth-Tormen (ST) mass function with that of Press-Schechter (PS) mass function. From the figure it is clear that the Sheth-Tormen mass function predicts a much larger volume filling of the universe. For example in atomic cooled models at \(z=3\), the ST mass function predicts 60% volume of the universe is filled with outflows (red dotted-dashed curve) compared to 35% with PS mass function (solid blue curve). Similarly for molecular cooled models at \(z=3\), the volume filling factors are \(F=0.65\) and 0.41 for ST and PS mass function respectively. Note that other physical properties do not change much when using different prescription for the halo formation rate. It is interesting to note that when we use the Sasaki formalism to calculate the formation rate of halos and include molecular cooled halos, the porosity decreases at \(z\sim 6\) before increasing again at \(z\sim 2\). In the Sasaki formalism, one has a survival probability for a halo collapsing at redshift \(z_c\) to survive till a later redshift \(z\). However when a halo is destroyed by merging to form a bigger halo, its outflow will survive, if it has gone far beyond the virial radius. Such outflows are not easy to take into account in a semi-analytical model such as ours, and we have neglected their contribution to the porosity. This can lead to a decrease in the volume filling factor. This effect is important when one calculates the volume filling factor of outflows from molecular cooled halos which collapse at high-\(z\), before reionization. On the other hand, models considering derivative of any mass function would over estimate the volume filling factor by considering outflows separately from halos after and before merging. Thus calculations which employ the formation rate using the Press-Schechter-Sasaki formalism and Sheth-Tormen-derivative bracket the expected behavior of volume filing factor, which is why we have considered both of them in the present work. Further, in we have tested the sensitivity of the porosity not only for various mass functions but also to different prescriptions for the formation rate of halos. It was shown that the filling factor as well as other physical quantities were reasonably insensitive to changes in the mass function. Thus we conclude that outflows can significantly volume fill the universe with metals, magnetic fields and cosmic rays. In passing we note that also got similar magnetic field strength of \(10^{-8}-10^{-12}\) Gauss in the IGM. However, their outflows never volume filled the universe as their simulation did not include small mass galaxies.
# Implications of outflow models {#sec_impl}
The outflow models discussed here has a number of interesting consequences, which we discuss below.
## Detectability of metals
In , we discussed the detectability of low density high temperature outflowing material in the presence of meta-galactic UV background through O vi and C iv absorption lines . In particular, we focused on absorption signatures that could arise from (i) the free wind, (ii) hot bubble and (iii) gas in the shell. The main difference in our current work is that, in the case of low mass halos the mass loading is higher and in the case of high mass halos the bubble temperature is lower when we include the CRs. Moreover, we are also considering magnetised outflowing material in the current work. Here we discuss the possible consequences without going into detailed radiative transport modelling. In , it was suggested that the low ion absorption seen through Mg ii, Na i and Ca ii transitions can not be produced in the free winds and could originate from multiphase structure likely to form inside the outflow due to (i) cold gas ejected from the galaxy along with the free wind, (ii) dense cloud formed from the Rayleigh-Taylor instability of the swept up shell or (iii) cold gas clumps formed due to cooling instabilities etc. We first ask, can the free wind driven by CR discussed here be detected in Mg ii absorption. In Fig.[\[fig_column\]](#fig_column){reference-type="ref" reference="fig_column"} we show the total integrated hydrogen column density produced by the free wind as a function of time for a halo of mass \(10^{11}\) M\(_\odot\). The integrated column density of total hydrogen inferred is very similar to what we had for models discussed in . All the discussions presented there is applicable here as well. In particular, to produce low ion absorption lines, we need multiphase structure with cold clumps originating from the above mentioned possibilities in addition to the flow structure discussed in this work. Next we consider the survivability of such cold clouds in the outflow. As discussed in , the evaporation time-scale goes as \(T^{-5/2}\). As can be seen in Fig. [\[fig_comp\]](#fig_comp){reference-type="ref" reference="fig_comp"}, the bubble temperature in our present models are typically factor 4 times smaller than for models considered in . This alone makes the evaporation time-scale \(\sim 32\) times longer than what we inferred for the pure thermally driven model without any feedback. Even when we consider the SNe feedback inclusion of CRs increases the evaporation time-scale by a factor 5. This suggests that survival of cold dense clumps in the outflow can be relatively easier in our models with CR driven flows. Further, in a magnetised outflow, magnetic fields can drape themselves around a gas cloud, and reduce evaporation of gas across the field even further.
## Thermal history of the IGM
We now explore the possible implications of the excess cosmic ray energy from outflows to the thermal history of the IGM. As noted earlier, observations of absorption lines in high redshift quasar spectra reveal that the temperature of the IGM at \(z=2-3\) is few times \(10^4\) K. This temperature is measured in the mildly overdense regions of the IGM with overdensity less than 10 as probed by the Lyman-\(\alpha\) forest. It is believed that the IGM temperature keeps some memory of the reionisation process at high-\(z\). At the end of the hydrogen reionisation the temperature of the IGM is set to \(1-3\times 10^4\) K due to photoionisation heating. In absence of any other heating mechanism after reionisation, the IGM temperature should drop adiabatically with the expansion of the universe as \((1+z)^2\). Thus in models where reionization of the universe occurred earlier than \(z=6\) (as expected from observations of quasar spectra) one would not be able to explain the observed IGM temperature at \(z=2-3\). We need an additional heating mechanism to explain this temperature floor. Several authors have suggested the late HeII to HeIII reionisation by the harder quasar spectrum as the reason for this excess temperature. Here we discuss if excess CR energy from outflows can heat the IGM at \(z=2\) to 3. Note that wherever galactic outflows percolate, they carry magnetic fields and cosmic rays in addition to metals. Thus, say at \(z=3\) for models assuming PS mass function, 50% of the universe is filled by the outflows from molecular cooled halos (35% for atomic cooled model) with \(\sim 10\) nano Gauss magnetic fields (optimistic model) and excess average cosmic ray energy density of \(10^{-4}\) eV/cm\(^{3}\). With ST mass function for atomic and molecular cooled models, respectively 65% and 41% of the universe is filled with outflows and similar amount of magnetic fields and excess average cosmic ray energy density. Note that in standard model of structure formation, the outflow affected regions are expected to collapse to form filamentary structures at latter stages that are likely to be detected as Lyman-\(\alpha\) forests in the quasar spectrum. The cosmic ray particles (protons) gyrate along the magnetic fields lines and generate Alfvén waves by a streaming instability. These waves damp out and transfer energy to the thermal gas. The rate of energy transfer to the gas via the Alfvén waves is given by \(|v_A.\nabla P_c|\) where \(v_A = B/\sqrt{4\pi \rho}\) is the Alfv́en velocity (\(\rho\) is the density of plasma). The modulus is taken as in this process CRs always lose energy. Note that for ionised gas with temperature \(T\) the thermal energy density is given by \(E_{th}=3n_Hk_B T\), \(n_H\) being the number density of hydrogen and \(k_B\) is the Boltzmann constant. Thus, time taken to deposit cosmic ray energy of the order of thermal energy is given by, \[t_{cr}= \dfrac {3nk_BT}{|v_A.\nabla P_c|}.\] Comparing this with the Hubble time \(t_H\) at a given epoch we get, \[\dfrac{t_{cr}}{t_H}= H(z)\dfrac {3nk_BT}{|v_A.\nabla P_c|}.\] Putting in numbers we find, \[\begin{aligned}
\dfrac{t_{cr}}{t_H} \approx 0.3 \left(\dfrac{h}{0.7} \right)^{4} \left(\dfrac{\Omega_m}{0.3} \right)^{1/2}\left(\dfrac{1+z}{4}\right)^6 \left(\dfrac{T_4}{B_{10 {\rm nG}}}\right) & \nonumber \\ \times \left(\dfrac{3\times 10^{-4} {\rm eV/cm}^3}{U_{cr}} \right) \left(\dfrac{L}{0.1~{\rm Mpc}} \right). & \label{eq-cr-heating} \end{aligned}\] Here, \(T_4\) is the temperature in the unit of \(10^4\) K and \(B_{10{\rm nG}}\) is the magnetic field in unit of 10 nano Gauss. Further we have normalised this ratio assuming a typical pressure gradient scale of a filament is \(L \sim 0.1\) Mpc and cosmic ray energy density \(U_{cr} = 3\times 10^{-4}\) eV/cm\(^3\). Hence, if the IGM at \(z=3\), has been polluted by outflows with \(B\) of 10 nano Gauss and \(U_{cr} = 3\times 10^{-4}\) eV/cm\(^3\), within 1/3rd of the Hubble time, the CRs would dissipate energy to the thermal gas via Alfvén waves, increasing IGM temperature to \(10^4\) K. For even lower redshift, the process is more efficient. In Fig. [\[fig_cr_heating\]](#fig_cr_heating){reference-type="ref" reference="fig_cr_heating"} we have plotted the volume average cosmic ray energy density available from our outflow models and compared it with the thermal energy of the ionised IGM of temperature \(10^4\) K. We have also shown thermal energy of the Lyman-\(\alpha\) forests with temperatures as measured by and. It is clear that the average cosmic ray energy density around \(z\sim 3\) (\(U_{cr} \sim 10^{-4}\) eV/cm\(^3\)) is much higher than the thermal energy density of the IGM with temperature \(10^4\) K, especially for molecular cooled models. The porosity weighted CR energy density would be even higher. In the mildly overdense regions of the IGM which produce Lyman-\(\alpha\) forest \(U_{cr}\) can be enhanced further by adiabatic compression to few times \(10^{-4}\) eV/cm\(^3\). Magnetic field will also be enhanced due to flux freezing to few tens of nano Gauss. Thus one can see from Eq. [\[eq-cr-heating\]](#eq-cr-heating){reference-type="ref" reference="eq-cr-heating"}, that CRs can transfer sufficient energy to the thermal gas within the Hubble time and potentially heat the IGM around \(z=2-4\) to a temperature of \(\sim 10^4\) K to erase the thermal signature of reionization. This conclusion would be more appropriate for molecular cooled models where outflows volume filled the universe early in time. Note that here we have considered CRs that are accelerated in the inner shocks only. However, we have not considered the CRs that can be generated in the terminal shock of the outflows as well as in the structure formation shocks. Simulations predict cosmic ray energy density of \(\sim 10^{-5}\) eV/cm\(^3\) that can arise from structure formation shocks with Mach number of few tens . Adding these sources of cosmic rays would lead to more efficient heating of the IGM by cosmic rays.
## IGM magnetic fields
As noted earlier, \(\gamma\)-ray observation of nearby blazars (with \(z \sim 0.1-0.2\)), have tentatively indicated a lower limit for magnetic fields in voids, \(B_{V}\sim 10^{-16}\) Gauss provided their coherence length \(l_B\) greater than about a Mpc. For fields with much smaller coherence scale, the lower limit increases as \(l_B^{-1/2}\) with decreasing \(l_B\). There fields in voids could be primordial arising from processes in the early universe. However, by such low redshifts, all our models predict a porosity of wind material greater than unity or volume filling factor greater than 0.6 for randomly distributed sources. This implies that outflows could have percolated even in the void regions by the present epoch. Further the porosity averaged value of the magnetic fields in our models are \(\sim\) nano Gauss, and the radius of the outflowing bubbles is of order Mpc. The coherence scale of the fields could be smaller than this radius due to turbulent tangling. And the field strength could be diluted further in outflows frozen into the expanding IGM in voids. Nevertheless, the field strength in our models are sufficiently large compared to \(B_V\), that it is highly likely that they satisfy the constraints set by the \(\gamma\)-ray observations. Thus magnetic fields inferred in the voids or in general IGM need not be of primordial origin. Rather they could be of astrophysical origin, generated/amplified in the star forming galaxies and transported from there to the IGM by the outflows created from the supernova explosions. One caveat is that large-scale clustering could result in a smaller volume filling of outflows in the void regions. It would be of interest to therefore consider our outflow models in more detail, in biased proto-void regions.
# Discussion and Conclusions {#sec_dc}
In the present work, we have explored how cosmic rays could aid in driving relatively cold outflows from high redshift galaxies, and resulting global consequences. The primary sources of energy for outflows are SNe in the high-\(z\) galaxies, and a fraction of this energy goes into cosmic rays. Outflows also cause a negative feedback on the star formation rate in a galaxy which can be especially large in low mass galaxies. We show that even with such suppression, low mass galaxies can still drive an outflow in the presence of cosmic rays. Cosmic rays are accelerated in the SNe shocks and produce cold free winds upto the inner shock and are then re-accelerated there. These cosmic rays add extra pressure to the thermal gas that can even drive a cold outflow. We find that outflows extend to tens of kpc in dwarf galaxies with halo mass \(\sim10^8\) M\(_\odot\) before merging with cosmological expansion, whereas massive galaxies with halo of masses \(10^{11}\) M\(_\odot\) have Mpc scale outflows. Cosmic ray driven outflows can go to a larger radii compare to a purely thermally driven outflow with same total energy input. This is because the cosmic ray pressure decreases less rapidly compared to the thermal pressure with adiabatic expansion (due to its softer equation of state) and comes to dominate outflow dynamics at large radii. We showed that these outflows could significantly volume fill the universe with metals, magnetic field and excess cosmic rays. In particular, at \(z=3\) for the ST mass function with molecular cooled halos, 65% of the universe is filled with the outflowing gas along with metals, magnetic fields of order \(10^{-8}\) G and excess cosmic ray energy density of \(\sim 10^{-4}\) eV/cm\(^3\). All our models predict a volume filling factor of order unity by \(z\sim 0\). Further, we explore three implications of such outflows. We argued that due to lower temperature and higher density of the free wind as well as the bubble gas, the metals spread by the outflow could be more easily detected as absorption line system. Low ionisation metals such as Na i or Mg ii are likely to be detected in gas clouds embedded the free wind, whereas others like C iv or O vi would come from the bubble or from pre-enriched IGM. Moreover, we showed that the excess energy density in cosmic rays from these outflows could be transferred efficiently enough to the thermal IGM gas in presence of magnetic fields. This can potentially erase the thermal history of reionisation and heat the IGM to a temperature of order \(10^4\) K as inferred from quasar spectra. Finally, we find that void magnetic fields of order \(10^{-16}\) G as inferred from blazar observation, need not be of primordial origin, but could be of astrophysical origin generated in star forming galaxies and transported by the cosmic ray driven winds to the voids. Our models which estimate the global impact of outflows on the IGM, at present do not take account of spatial clustering of galaxies. In order to have more realistic scenario, it would be of interests to put the cosmic ray driven outflows in numerical simulation of structure formation. | {'timestamp': '2018-02-01T02:12:13', 'yymm': '1706', 'arxiv_id': '1706.01890', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01890'} |
# Introduction | {'timestamp': '2017-06-07T02:04:20', 'yymm': '1706', 'arxiv_id': '1706.01647', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01647'} |
# Introduction
Glitches are sudden jumps of amplitude \(\Delta\Omega\) in the angular velocity \(\Omega\) of some isolated pulsars with fractional increases that span the range \(\Delta\Omega/\Omega \sim 10^{-11} \div 10^{-5}\) in otherwise steadily spinning-down neutron stars (NSs). According to the Jodrell Bank Glitch Catalog[^1], more than four hundred glitches have been observed, in more than a hundred objects including rotational-powered pulsars, magnetars and millisecond pulsars. Although there is some evidence that small glitches in the young Crab pulsar are caused by star quakes, the widely accepted explanation for large glitches is based on the presence of a superfluid component (made of paired neutrons) that can rotate differentially with respect to the non-superfluid rest of the NS. The superfluid stores extra angular momentum that can be abruptly exchanged with the normal component, resulting in a little decrease of the observed rotational period of the pulsar (for a review on glitch models see ). A rotating superfluid is permeated by many vortex lines (of the order \(\sim 10^{19}\) for the Vela pulsar) of quantized circulation. The exchange of angular momentum is mediated by these vortices that induce a mutual friction between the superfluid and normal components. The two physical ingredients that currently challenge the description of pulsar glitches, both at the mesoscopic scale of vortex lines as well as at the macroscopic stellar scale, are entrainment and pinning.\
\
*Entrainment-* A key feature of superfluidity is the possibility for the normal and superfluid components to flow independently although they compenetrate. The formalization of this multi-fluid problem in the interior of NSs is based on the seminal works on the hydrodynamics of He-II that extended the early two-fluid models of Tisza and Landau. Later, with the aim to study superfluid solutions, Andreev and Bashkin developed a three-fluid framework where two superflows and a normal flow exist simultaneously. They found that entrainment (a non-dissipative effect that couples the two species) can be described in the frame of reference comoving with the normal flow as a "mass-matrix" which relates the momentum of one constituent to the kinematic velocities of both constituents (see also for modern formalism with applications to the hydrodynamics of NS interiors). Thus entrainment is also discussed, both in the core and in crust, in terms of effective masses (which definition depend on the reference frame used). Dripped neutrons in NS crusts are analogous to electrons in ordinary metals and their properties can be determined by the band theory of solids: only neutrons in the conduction band can move throughout the crust. Despite the absence of viscous drag, the crust resists the neutron current due to Bragg scattering. This is a non-local and non-dissipative effect: neutrons are scattered by individual clusters and can interfere constructively or destructively. In the latter case neutrons cannot propagate, as if they were very massive. The scenario is different in the core, where neutrons and protons can only be locally scattered by surrounding nucleons. The resulting low effective mass of neutrons in the core can be interpreted in terms of a back-flow of nucleons. This local effect also exists in the crust but it is negligible when compared to Bragg scattering.\
Entrainment challenges our understanding of pulsar glitches. In particular Vela glitches have been thought to arise from the superfluid in neutron-star crust. However superfluid can be strongly entrained by the crust and this results in a decreased mobility of the crustal superfluid that cannot carry enough angular momentum to explain large glitches. In our numerical examples we use the effective neutron masses given in and in for the core and the crust respectively.\
\
*Pinning-* Pinning is the mechanism that allows to build up the angular momentum reservoir: when vortices are blocked the normal component slows down due to radiation losses while the pinned superfluid maintains its state of motion, lagging behind and storing angular momentum which can then be released in a glitch.\
Pinning is a well known phenomenon that occurs in type-II superconductors, where the flux-tubes can pin to defects of the crystalline structure. Something analogous happens in NSs: when vortex lines are immersed into a background current (of superfluid neutrons) an hydrodynamical lift develops and the vortices tend to drift outside the star (vortex-creep). Strong pinning allows to support higher currents before the dissipative effects generated by the motion of vortices enter the game; this in turns imply that a bigger angular momentum reservoir can be stored into the current of superfluid neutrons. The strength of pinning is described by a mesoscopic quantity \(f_P\) that defines the depinning threshold: it is the module of the maximum lift force on vortices that the pinning centers can sustain. At a microscopic scale it is possible to calculate the *individual pinning energies* of a single vortex with a single pinning center but to find the maximal non-dissipative current of neutrons that pinning can sustain we need the *average pinning force* of vortices passing trough many pinning sites. Typically in NSs pinning refers to the interaction of vortices with the crustal lattice and pinning centers are the nuclei of the inner-crust. In this work we will not consider the further possibility of pinning between vortex lines and flux-tubes of the proton superconductor, even if this suggests interesting scenarios. However a vortex-fluxtube pinning profile, that is a consequence of entrainment in the core, can be added to that of crustal pinning we use, without any change in our model.\
\
In this presentation we expand the discussion of the model proposed in . This model, despite its simplicity, can be used to test quantitatively the widespread assumption of a stable vortex array in superfluid NSs. In fact, vortex-creep models (since the seminal work of ) do not account for entrainment or stratification. Similarly, global models based on the two-fluid formalism have either neglected entrainment or have been studied under the assumption of uniform entrainment and rigid rotation of the neutron superfluid.
# Application to pulsar glitches: axially symmetric model
The multi-fluid formalism used to describe the superfluid hydrodynamics in NS interiors is well established and in principle enables us to model the glitch and the subsequent relaxation. However in a three-dimensional hydrodynamical simulation it is not simple to take into account consistently for stratification, pinning forces and entrainment. We show how to implement consistently these ingredients within a simplified geometry, as a first natural approximation to the full problem. In this way we rule out a priori turbulent motion and its consequences on the vortex-mediated mutual friction: this is a limit of the model to be kept in mind. The model is already detailed in, in this section we discuss its features and assumptions:
- The star spins around the z-axis [^2] (there is no precession); since we study non-millisecond pulsars, we consider that the equilibrium structure is given by solving the TOV equations. As sketched in Fig 1, the superfluid is contained into a spherical volume of radius \(R_d\) (the radius at which the neutron drip starts, corresponding to a density of \(\sim 4\times10^{11}\,\)g cm\(^{-3}\)).
- In the absence of a layer of normal matter between the core and the crust, the vortex cores can pass continuously through this smooth transition. Leaving out the problem of the unknown composition of the inner core, we consider that the superfluid component (the n-component) is threaded by straight vortices parallel to \(\hat{\bm{e}}_z\). However it is equally simple to consider superfluidity only inside a spherical shell of given radii. The scenario of continuous vortex lines was already suggested in as an alternative to the scenario of distinct vorticity in the crust and in the core.
- The rigid crust and the charged component (i.e. electrons, protons and bounded neutrons, that we call p-component) are locked into the strong magnetic field of the neutron star on short timescales . Electromagnetic losses produce a constant braking torque \(-I \dot\Omega_{\infty}{\bm{e}}_z\) that acts on the p-component, where \(I\) is the total moment of inertia of the star and \(\dot\Omega_{\infty}\) the absolute value of the spin-down rate over a long period of time. The velocity field of this rigid p-component is \(\bm{v}_p = x \Omega_p \hat{\bm{e}}_\theta\), which gives the observed pulsar period \(2\pi \Omega_p^{-1}\).
- The presence of vortices induces a mutual friction between the n-component and the p-component: scattering of particles (or quasiparticles) off vortex cores gives rise to mesoscopic drag force \(\bm f_D\).
- Vortex lines can pin to nuclei in the solid inner crust. The force per unit length \(f_p(\rho)\) expresses the strength of the vortex-lattice interaction.
We are thus far from describing the realistic hydrodynamical problem of a NS interior; superfluid turbulence (a well known phenomenon in He-II where it gives rise to a complex tangle of vortex lines ) is likely to develop in some regime also in NSs, but its effect on macroscopic hydrodynamics and its relevance for glitch models are still debated. In particular the tension of a single line is much smaller than the hydrodynamical force acting on it, which suggests bending of vortices, such that the core is susceptible to become turbulent. However this argument does not account for the stabilizing effect of rotation: the increased effective tension of a bundle of vortices as compared to a single vortex line resists bending and the array of vortex lines could remain parallel to the rotation axis. This can be seen as a generalization of the Taylor-Proudman theorem to axisymmetric and spinnig-down frictionless fluids.\
In the zero temperature limit the microphysical quantities of interest, as the the superfluid fraction \(x_n(\rho)\) and the dimensionless effective mass \(m^*(\rho)= 1-\epsilon_n(\rho)\) of entrained neutrons, can be easily parametrized as functions of the spherical radius \(r\) after solving the TOV equations for a given EOS. We impose a Newtonian dynamics on the general relativistic stellar structure, thus the total moment of inertia is given by \[I \, = \, \frac{8 \pi}{3}\int_0^R \! r^4 \rho(r) \,\,dr \,.\] General relativistic corrections to this quantity can be computed using the Hartle slow-rotation approximation, where the background structure is still given by the TOV equations and are expected to be large (see e.g. for a recent description of this approximation). We further discuss this point in the concluding remarks.\
\
*Conservation of vorticity-* Following the Newtonian formalism outlined in, the momentum per particle of the superfluid is a linear superposition of the local kinematic velocities of both components: \[\bm{p}_n/m_n \,=\, \left(1-\epsilon_n\right) \bm{v}_n\, +\,\epsilon_n \bm{v}_p \,,\] where \(m_n\) is the neutron bare mass. At the microscopic scale \(\bm{p}_n\) is irrotational but here we consider it as averaged over a region that contains many vortex lines. In our axially symmetric configuration the azimuthal component of the momentum turns out to be columnar (it has no dependence on \(z\)). It is thus convenient to introduce an auxiliary angular velocity \(\Omega_v(x)\) defined by \[\bm{p}_n(\bm{x})\cdot \hat{\bm{e}}_\theta = m_n\,x \,\Omega_v(x) \,\,.\] On the other hand, the angular velocity of the n-component is non-columnar due to the (spherically symmetric) density dependence of the entrainment parameter \(\epsilon_n(r)\). Exchange of angular momentum between the superfluid and the p-component is possible only if we allow for radial motion of vortices. Let \(v_L^x\) be the radial component of the local mean velocity of vortex lines (the velocity of a vortex is normal to the line), then a simple argument early proposed in gives \[\label{eq:1} \dot\Omega_v(x,t) =-\left( 2\Omega_v(x,t) + x\partial_x \Omega_v(x,t) \right)\frac{v_L^x(x,t)}{x} \,\,.\] This equation expresses a conservation law for vortex lines flowing trough a cylindrical shell of radius \(x\). The radial velocity \(v_L^x(x,t)\) needs to be expressed in terms of the variables \(\Omega_v(x,t)\) and \(\Omega_p(t)\) by solving an explicit model of vortex dynamics. In a simple example is proposed: we construct a particular form of the friction functional \(\mathcal{B}\) defined as \[v_L^x(x) \,\, = \,\, x \,\, \mathcal{B}[\Omega_v,\Omega_p,x] \,\, (\Omega_v(x)-\Omega_p) \.\] The explicit form of \(\mathcal{B}\) depends on how vortex dynamics is modeled; correlated motion of many lines, pinning and finite tension of macroscopically long vortices make the explicit construction of this functional a non trivial task.\
\
*Total angular momentum balance-* The isolated pulsar dynamics is slowly driven by angular momentum losses due to radiation emission, namely \[\int\!d^3x\,\rho(r)\,\left[ \, x_n(r) \, \bm{x} \times \dot{\bm{v}}_n + (1-x_n(r)) \, x^2\, \dot{\Omega}_p\, \hat{\bm{e}}_z \, \right] \,=\,-I\,\dot\Omega_{\infty}\hat{\bm{e}}_z \,.\] Expressing the non-columnar velocity of the n-component in terms of the angular velocities, the angular momentum balance takes the form \[\label{eq:2} I_p \, \dot{\Omega}_p \, + \, I_v \, \langle \, \dot{\Omega}_v \, \rangle \, = \,-I \, \dot{\Omega}_\infty \,,\] where \(\Omega_v\) is averaged over the cylindrical radius \(x\), namely \[\langle \, \dot{\Omega}_v \, \rangle \, = \, \frac{1}{I_v} \int_0^{R_d} \! dI_v(x) \,\, \dot{\Omega}_v(x,t) \,.\] The partial moments of inertia are \(I_p = I-I_v\) and \(I_v\), that is also the normalization factor of the measure \(dI_v\). The physical constraint on entrainment parameter \(\epsilon_n < 1-x_n\) guarantees the positivity of this quantity, namely \(I_v<I\). The explicit Newtonian form of \(dI_v\) is \[\frac{dI_v(x)}{dx} \,=\, 4\,\pi\,x^3 \int_{0}^{z(x)} \! dz \,\, \frac{\rho_n(r)}{m^*(r)} \,,\] with \(z(x)\) the height of vortices passing through \(x\). The appearance of the effective mass at the denominator is physically reasonable: the effect of entrainment is to redefine the mass of neutrons (i.e. \(\bm p_n = m^*\bm v_n\) in the frame corotating with the p-component) that rescales \(\rho_n\).\
At the steady-state (i.e when \(\dot \Omega_p=\dot \Omega_v\) for all the \(x\) values ), Eq [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"} assures that the star responds to the braking torque as a whole with a spin-down rate of \(-\dot{\Omega}_\infty\). Thus, provided that the braking torque itself is not evolving, the value of \(\dot{\Omega}_\infty\) coincides with the absolute value of the observed mean spin-down rate over a long period of time (even if in this period the pulsar glitched several times). It is also useful to introduce the lag \(\omega\), defined as \[\omega(x,t) \, = \, \Omega_v(x,t)-\Omega_p (t) \,.\] At the steady state the lag is constant in time and strictly positive for all the \(x\) values.\
\
*Friction functional-* Vortex lines experience a viscous drag force per unit length \(\bm{f}_D\) and an hydrodynamical lift (the Magnus force) \(\bm{f}_M\) expressed by \[\begin{aligned}
\bm{f}_D &=-\eta(r) ( \bm{v}_L(\bm{x})-\bm{v}_p(\bm{x}) ) \,, \\ \bm{f}_M &=\rho_n(r) \bm{\kappa}(\bm{x}) \times ( \bm{v}_L(\bm{x})-\bm{v}_n(\bm{x}) ) \,. \end{aligned}\] As discussed in the vector \(\bm{\kappa}\) is given by \[\bm{\kappa}(\bm{x}) \,=\, \kappa \,\nabla \times\bm{p_n} \, / \, |\nabla \times \bm{p_n}| \,.\] Within our working assumptions, this field is constant and parallel to \(\hat{\bm{e}}_z\). To implement this infinite rigidity of vortex lines, the equation of motion of vortices is modeled as \[\bm{F}_{tot} \,=\, \int_L \! dl(\bm{x}) \,[ \bm{f}_M(\bm{x})+\bm{f}_D(\bm{x}) ] \,=\,0 \,,\] where the integration is performed over the vortex line. Solving the above equation for the radial component of the vortex velocity gives that the friction functional \(\mathcal{B}\) is just a function \(\mathcal{B}^x\) of the cylindrical radius \[\label{eq:free-velocity} v_L^x \, = \, x \, \mathcal{B}^x(x) \,\omega(x,t) \,.\] At this level we still have to account for pinning: vortex lines that are perfectly pinned are blocked and forced to corotate with the p-component, until the Magnus force induced to vortex lines by the background flow of neutrons overcomes the depinning critical value. This threshold effect introduces explicit and non-smooth dependence of \(\mathcal B\) on the dynamical variables. The depinning condition can be found by imposing that \[|\bm{F}_{tot}|_{\bm v_L =\bm v_p} \,=\, \int_L \! dl \,\frac{\kappa\,\rho_n}{m*} \,x\,\omega \,=\, \int_L \! dl \,f_P\] and a straightforward calculation gives that the second equality is realized when the dynamical lag \(\omega\) reaches the value \[\label{eq:clag} \omega_{cr}(x) \, = \, \frac{ \int_{0}^{z(x)} \! dz \,\, f_P(r) }{\kappa \, x\, \int_{0}^{z(x)} \! dz \,\,\rho_n(r)/m^*(r) } \,.\] This function of the cylindrical radius (sketched in Fig 2 for a typical case) reduces to the critical lag for depinning studied in and in the limit of no entrainment. At the level of equation of motion for rigid lines, pinning could be modeled by introducing a local drag force with a non-constant \(\eta\) parameter that depends on the dynamical lag \(\omega\) as \[\begin{aligned}
\eta(\bm x,\omega) \, \gg \, \kappa \rho_n(r) \qquad &\text{if } \quad |\omega(x)| \, < \, \omega_{cr}(x) \, \,, \\ \eta(\bm x,\omega) \, \ll \, \kappa \rho_n(r) \qquad &\text{if } \quad |\omega(x)| \, \simeq \, \omega_{cr}(x) \, \. \end{aligned}\] As an order of magnitude estimate, if \(\eta \sim 10^{21}\,\)s\(\,\)cm/g in the crust, the \(\mathcal B^x\) of Eq [\[eq:free-velocity\]](#eq:free-velocity){reference-type="ref" reference="eq:free-velocity"} ranges from \(\mathcal{B}^x(x=0) \sim 10^{-9}\) to \(\mathcal{B}^x(x=R_d) \sim 10^{-12}\), thus ensuring that the vortices are expelled on timescales of many years.\
However, since we are just interested to estimate the angular momentum reservoir, we can rely to a simpler possibility that is more similar to the early vortex-creep model, where the effect of the dynamical lag \(\omega\) against pinning is to modulate the typical velocity of free vortices given in Eq [\[eq:free-velocity\]](#eq:free-velocity){reference-type="ref" reference="eq:free-velocity"}. This is equivalent to approximate the complete (and unknown) friction functional as \[\mathcal{B}[\Omega_v,\Omega_p,x]\,\, \approx \,\,Y[\omega, x]\, \,\mathcal{B}^x(x) \,.\] From the macroscopic point of view, the fraction \(Y \in [0,1]\) can be equally interpreted as a parameter that slows the expulsion of vortex lines or the fraction of (unpinned) vortex lines that are free to move under the combined action of \(\bm{f}_M\) and \(\bm{f}_D\). By considering a large population of vortices, \(Y\) can also be seen as a local depinning probability. With macroscopically straight vortex lines, the simplest way to encode pinning at the macroscopic scale is thus to construct \(Y\) in such a way that \[\begin{aligned}
Y[\omega, x] \, \approx 0 \qquad &\text{if } \quad |\omega(x)| \, < \, \omega_{cr}(x) \, \,, \\ Y[\omega, x] \, \approx 1 \qquad &\text{if } \quad |\omega(x)| \, \simeq \, \omega_{cr}(x) \, \. \end{aligned}\] Finite temperature, quantum effects and finite vortex tension take part in smoothing the step-like form of \(Y\). As an explicit example, the original vortex-creep model in the regime \(0 < \omega < \tilde{\omega}_{cr}\) is realized by \[Y[\omega, x] \, \approx \, \exp \left( \frac{\omega(x)-\tilde{\omega}_{cr}(x)}{\alpha \,\tilde{\omega}_{cr}(x)} \right) \,,\] where \(\alpha\) is a dimensionless temperature that tunes the local rigidity of the pinning and \(\tilde{\omega}_{cr}(x)\) is an opportune critical lag. In the limit \(\alpha \rightarrow 0\) the vortex-creep prescription results into a perfect pinning prescription, namely \(Y=\theta(\tilde{\omega}_{cr}(x)-|\omega(x)| )\) is the unit step function.\
\
*Improving the friction term-* Vortices in an hydrodynamical model only serve to construct a particular form of the mutual friction in terms of the dynamical variables of the problem (in our global model we shifted this problem on \(\mathcal{B}\)). A glitch can then be triggered by perturbing \(\mathcal{B}\) or, in order to mimic the sudden unpinning of many vortices, by suddenly increasing the \(Y\) fraction in some interval of \(x\). It is thus important to find good prescriptions for the friction functional, even in global or averaged models like the one discussed here. Other contributions to the vortex dynamics (that could have a significant impact on mutual friction) arise from the finite tension force of vortices and their mutual interaction. Self-interaction forces that complicate further the picture are present for curved lines. Moreover the radiation of phonons during the pinning and depinning of a vortex line and proximity effect when two lines are close are important at a microscopic scale: these precesses could provide the physics needed to explain the claimed self-organized criticality of glitching pulsars. Further work on these lines is needed to understand how these issues modify mutual friction.
# Constraints from large glitches
For any given lag \(\omega\), the corresponding reservoir of angular momentum (i.e. the difference between the total angular momentum and that of the star in perfect corotation) is \[L \,-\, I\,\Omega_p \,=\, I_v \, \langle \omega \rangle \,.\] The spin-up phase of glitches is so fast that the braking torque carries out a negligible amount of momentum; the total angular momentum is thus conserved during a glitch \[\dot{L} \,=\, I_v \, \langle\dot{\omega} \rangle \,+\, I \, \dot{\Omega}_p \,=\, 0 \,.\] Assume now to trigger a *maximal* glitch (a glitch that by definition empties the whole reservoir) at a certain initial time: there will be a subsequent time at which the star is in a state of instantaneous average corotation, namely \(\langle \omega \rangle =0\). This situation can be realized during the spin up phase or just after an overshoot, a transient period during which \(\langle \omega \rangle<0\). Integration of the previous equation between these two times gives \[I_v \, \langle \Delta \omega \rangle \, + \, I \Delta \Omega_p \, = \, 0 \,,\] where \(\Delta\) indicates the difference between a quantity at the particular time when corotation is instantaneously realized and the same quantity at the triggering time.\
\
*Entrainment independent maximum glitch amplitude-* In our model the maximum sustainable lag is the critical depinning lag \(\omega_{cr}\) (the actual maximum sustainable lag is likely to be less than this quantity, because of thermal effects and dynamical creep) so that we can set an absolute upper limit to the glitch amplitude by imposing \(\Delta \omega \, = \,-\omega_{cr}\): \[\label{eq:max-gl} \Delta \Omega_{max} \, = \, \frac{I_v}{I} \, \langle\omega_{cr} \rangle \.\] Using the explicit forms of \(dI_v\) and \(\omega_{cr}\), it is trivial to show that \(\Delta \Omega_{max}\) depends only on the EOS and on the pinning forces: at the Newtonian level entrainment does not affect this quantity (this ceases to be the case when relativistic corrections are taken into account). Thus, by fixing these two microphysical inputs, it is possible to study how \(\Delta \Omega_{max}\) changes with the total mass of the NS. This is shown in Fig 3, together with the maximum glitch amplitude \(\Delta\Omega\) recorded in some pulsars; we thus find an upper limit to the mass of each pulsar by imposing \(\Delta\Omega=\Delta\Omega_{max}(M)\). As a general trend, a stiff EOS like GM1 gives less stringent upper mass bounds (e.g. \(M_{Vela}<1.4\) with SLy and \(M_{Vela}<1.8\, M_{\odot}\) with GM1). Future observations of even larger glitch amplitudes can only lower this upper limit to the mass, leading to stronger constraints. Note also that the possible presence of turbulence in the crust could reduce the maximum glitch amplitude, thus lowering the curves of Fig 3 and leading to stronger constraints: it seems reasonable to expect that the increase in the vortex length density due to the development of complex vortex tangles weakens the otherwise coherent pinning of the vortex array. Finally we remark that, as long as only crustal pinning is considered, \(\Delta \Omega_{max}\) turns out to be the same for vortex lines that stop at the crust-core interface or pass through the core.\
\
*Constraints imposed by the reservoir dynamics-* It is possible to try to lower the upper bound of the mass by using also observed timing properties other than the largest glitch. Starting from corotation, forward integration in time of Eqs [\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"} and [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"} allows to follow the evolution of the quantity \[\label{eq:dynamical-max-gl} \Delta\Omega_{m}(t) \, = \, \frac{I_v}{I} \, \langle \omega(x,t) \rangle \,\] Equation [\[eq:dynamical-max-gl\]](#eq:dynamical-max-gl){reference-type="ref" reference="eq:dynamical-max-gl"} gives the amplitude of a maximal glitch, triggered at time \(t\) after the emptying of the reservoir because of an initial large event. This results in a correlation between the amplitude of maximal glitches and the waiting time between them. This of course ceases to be interesting if there are no pulsars that sporadically empty the reservoir: in this case the sequence of amplitudes and waiting times shows no correlations between events, which is certainly true for glitches of small amplitude that comprise the major part of the whole sample among the pulsar population. However J0537-6910 and Vela have been proven to be exceptional objects: systematic study of glitches in individual pulsars seem to indicate that the distribution of waiting times is Poissonian while the distribution of glitch amplitudes is a power law (even though the value of the exponent is not universal across the pulsar population), but Vela and J0537-6910 showed evidence for quasi-periodicity. The main difficulty behind this and similar analysis is the paucity of data in single objects.\
In the following we present a more speculative set of ideas: instead of solving dynamical equations with the particular rotational parameters for each pulsar, we use the simple unified model sketched in Fig 2 and defined as \[\label{eq:unified-model} \Delta\Omega_{m}(\omega^*) \, = \, \frac{I_v}{I} \, \langle \, \min{(\omega_{cr}(x),\omega^*)} \, \rangle \.\] For each pulsar we measure time in terms of a nominal lag \(\omega^* = t\, \dot\Omega_{\infty}\). The curves defined by Eq [\[eq:unified-model\]](#eq:unified-model){reference-type="ref" reference="eq:unified-model"} are shown in Fig 4 for different masses. On top of this, we plot the single glitches of some pulsars (among which Vela and J0537-6910) versus waiting times to the previous glitch multiplied by the absolute value of the spin-down rate \(-\dot\Omega\) the object. The largest glitch of J0537-6910 (the only pulsar that display correlation between the amplitude and the time to the next glitch ) is also the first one, so we plotted it at the end of the \(\omega^*\) axis, where \(\Delta\Omega_{m}(\omega^*)\) assumes the value given by Eq [\[eq:max-gl\]](#eq:max-gl){reference-type="ref" reference="eq:max-gl"}. It is interesting to note that is always the largest glitch that touches the curve corresponding to the smaller mass, even if the data are scattered. Note that for each glitch we used the nominal lag corresponding to the time since the previous event, without taking into account the amplitude of this event, even if it is unlikely that small glitches are maximal. This only serves as a clue of the fact that the largest glitch recorded in a pulsar can constrain its mass: at this level no conclusions can be taken on the actual (possibly time-correlated) occurrence of large glitches in pulsars.\
\
*A remark on the unified model-* We stress that when we simulate the full dynamical equations, starting from perfect corotation at \(t=0\), the lag evolves in a fashion very similar to the lag sketched in Fig 2: the horizontal plateau (at height \(\omega^*\)) rises until it locally approaches \(\omega_{cr}(x)\), where sub-critical vortex creep shapes the actual lag around \(\omega_{cr}\). Thus it is not a bad approximation to recast the dynamical evolution of the lag into the form \[\omega(x,t) \, \approx \, \min{\left(\, \omega_{cr}(x) \,, \, \omega^*(t)\,\right)} \,.\] However this can only serve to *define* \(\omega^*(t)\), a function of time that can be easily extrapolated from \(\omega(x,t)\) once the actual evolution is known from the forward integration in time of the dynamical equations (e.g. \(\omega^*(t)=\omega(x_m,t)\) where \(x_m\) is the cylindrical radius corresponding to the peak of \(\omega_{cr}\)). On the other hand, in the previous section we assumed \(\omega^*(t) = \dot\Omega_{\infty}t\). This choice is convenient since it is possible treat different pulsars within a unified prescription; in this way there is no need to solve the dynamical equations with the specific rotational parameters \(\Omega_p(0)=\Omega\) and \(\dot \Omega_{\infty} =-\dot \Omega\) for every different object. However there are no rigorous arguments for doing this, except for the fact that the slow dynamics of \(\omega\) toward the steady state is driven by \(\dot \Omega_{\infty}\). Moreover this unified prescription based on a linear relation between the actual time and the nominal lag \(\omega^*\) was already used (in a different context and with a different idealization of the lag dynamics) in some previous works that use the same form of the critical profile given in Eq [\[eq:clag\]](#eq:clag){reference-type="ref" reference="eq:clag"}, but without entrainment.
# Effect of entrainment
We switch off the entrainment effect setting \(m^*(r)=1\) everywhere and repeat the integrations needed to calculate the macroscopic input quantities of model. In this limit \(\omega \rightarrow \Omega_n-\Omega_p\), \(I_v \rightarrow I_n\), \(I_p \rightarrow I-I_n\). The curves \(\Delta\Omega_{m}(\omega^*)\) of the unified model (Eq [\[eq:unified-model\]](#eq:unified-model){reference-type="ref" reference="eq:unified-model"}) are found to be considerably steeper than their "entrained" counterparts, as shown in Fig 5. This is not surprising since the strong crustal entrainment sensibly increase the peak of the critical lag which in turn implies that the plateau at the value \(\Delta\Omega_{max}\) is reached for smaller values of the nominal lag. This is in part a drawback of the assumed form of \(\omega^*(t)\) that in general is not simply linear in time. Nonetheless the result seems physically reasonable as it is more difficult for strongly entrained superfluid neutrons to lag behind the crust. However a true comparative study of entrained and non-entrained dynamics should be carried by solving the dynamical equations [\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"} and [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"}. Despite this caveat this preliminary analysis is encouraging: large glitches of Vela and J0537-6910 seem to have been triggered around \(\omega^* \sim 10^{-2}\,\)rad/s, as can be seen in Fig 4. For this value of the nominal lag the difference between the case with entrainment and that without entrainment results in at most a \(\sim 10\%\) uncertainty on the mass value.
# Conclusions
We presented a global model for pulsar rotational dynamics that accounts consistently for the layered structure of the star, the differential rotation of the superfluid and the presence of density-dependent entrainment. The Eqs [\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"} and [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"} describe the exchange of angular momentum between two effective components fo the star, characterized by the partial moments of inertia \(I_v\) and \(I_p\). All the complex behavior of the rearrangement of the vortex configuration that should appear in Eq [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"} is hidden into the functional \(\mathcal{B}\). Even without solving explicitly the dynamical equations, it is possible to draw some general quantitative conclusions; in particular the calculation of the maximum glitch amplitude (that is entrainment independent and is unaffected by the possible presence of a normal layer at the crust-core interface) seems to open two possibilities:
1. Test pinning newly calculated forces using observations of large glitches.
2. Set an upper limit to the mass of some isolated pulsars (once the EOS and the pinning force profile have been fixed).
Finally we sketched a more speculative set of ideas that could be useful to refine the upper bound to the mass of large glitchers: while the absolute upper limit is fixed by the observation of the largest glitch only, to improve this constraint we have to use some additional information about the temporal occurrence of glitches in a pulsar.\
A final remark is due: we used the TOV equations for the structure of the star, but the classical moments of inertia. Fortunately it is quite simple to introduce general relativistic corrections for slow rotation to \(I\) and \(I_v\). Preliminary study indicates that (when vortex lines extend into the core, as in this presentation) the quantity \(I_v/I\) is nearly unaffected by relativistic corrections. However general relativity does not affect only the moments of inertia but also the measure \(dI_v\) and the critical lag; since in this Newtonian model we used the paraxial vortex configuration of a rigidly rotating body, the corrections to the cylindrical profiles could be estimated by resolving the stationary configuration of vortex lines for a slowly and rigidly rotating neutron star. Despite all possible caveats and simplifications that are present in our model, the message we want to vehicle is that also the study of single glitches (and in particular the largest glitch recorded in each pulsar) can robustly constrain the mass of isolated neutron stars: large glitches can provide a test for the consistency of glitch theory that is alternative to the studies based on pulsar activity. | {'timestamp': '2017-06-07T02:07:19', 'yymm': '1706', 'arxiv_id': '1706.01794', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01794'} |
null | null |
null | null |
# Introduction {#sec:intro}
Inflation describes a phase of accelerated expansion in the early Universe, during which vacuum quantum fluctuations of the gravitational and matter fields were amplified to cosmological perturbations . These primordial fluctuations later seeded the cosmic microwave background (CMB) anisotropies and the large-scale structure of our Universe. At present, the full set of observations can be accounted for in a minimal setup, where inflation is driven by a single scalar inflaton field with canonical kinetic term, minimally coupled to gravity, and evolving in a flat potential in the slow-roll regime . In most of these models, the potential becomes steeper as inflation proceeds and inflation eventually stops when the potential is not flat enough to support it. However, several mechanisms can lead to a premature end of inflation (PEI). In hybrid inflation for instance, an auxiliary field \(\chi\) is coupled to the inflaton field \(\phi\) through a term \(\propto \phi^2 \chi^2\), so that the effective mass of \(\chi\) is \(\phi\)-dependent. At early time during inflation, \(\chi\) has a large mass compared to the Hubble scale and it is a mere spectator field. As \(\phi\) rolls down its potential, the mass squared of \(\chi\) decreases and eventually becomes negative. This triggers a tachyonic instability that quickly terminates inflation (see however for cases where this "waterfall" phase extends over several \(e\)-folds). Recently, the geometrical destabilization (GD) of inflation has offered a new mechanism that can prematurely terminate inflation. The GD is the phenomenon by which the field space curvature can dominate forces originating from the potential and destabilize inflationary trajectories. Similarly to the well-known eta problem (see for instance ), it is a manifestation of the sensitivity of inflation to the physics near the Planck scale. It thus represents both a universal challenge and an opportunity to learn about the highest energy scales from cosmological observations. In the following, we make use of a minimal realization of the GD outlined in . A well-motivated possible outcome of this phenomenon is that it terminates inflation abruptly, much earlier than what slow-roll violation would have yielded. When this premature end of inflation occurs, the location of the observational window along the inflationary effective single-field potential is modified. This means that the part of the potential that the inflaton field is exploring when the modes of astrophysical interest today crossed the Hubble radius changes, and so do the predictions of a given model of inflation. In the present work, we confront the predictions of the models with a premature end of inflation with data, using Bayesian techniques. In practice, we consider various classes of prototypical models that are differently affected by such a phenomenon, which allow us to discuss most possible effects resulting from a change in the end of inflation time. Since the location of the observational window is also affected by the expansion history of reheating, which determines how far the observational window is from the end of inflation, we carefully incorporate this epoch and describe its degeneracies with the end of inflation location. The paper is organized as follows. In , we introduce the physical setup and the method used in this work. The geometrical destabilization is presented in , the role played by reheating in , the Bayesian model comparison approach in and the prototypical models we use in . In , we analyze our results, both at the level of posterior distributions on the parameters of the models and regarding their relative Bayesian evidences. Our main conclusions are summarized in .
# Setup and method {#sec:method}
## Geometrical destabilization and premature end of inflation {#sec:GD}
The geometrical destabilization of inflation is a generic phenomenon that potentially affects all realistic models embedded in high-energy physics. For simplicity, we only make use in this paper of a minimal realization of it through the following phenomenological two-field Lagrangian: \[\begin{aligned} \mathcal{L} =-\left(1+2\frac{\chi^2}{M_{\mathcal{R}}^2}\right)\frac{\left(\partial\phi\right)^2}{2}-V\left(\phi\right)-\frac{\left(\partial\chi\right)^2}{2}-\frac{m_h^2}{2}\chi^2\,. \label{minimal} \end{aligned}\] Let us first consider the situation in which the interaction \(\propto (\partial \phi)^2\chi^2\) is absent. The Lagrangian ([\[minimal\]](#minimal){reference-type="ref" reference="minimal"}) then describes an inflaton field \(\phi\), slowly rolling down its potential \(V(\phi)\), with an additional scalar field \(\chi\) which is heavy during inflation for \(m_h^2 \gg H^2\) (where \(H\equiv \dot{a}/a\) is the Hubble scale, \(a\) is the scale factor and a dot denotes derivation with respect to cosmic time), so that it is anchored at the bottom of the inflationary valley at \(\chi=0\). Let us now assess the impact of the kinetic coupling \(-(\partial \phi)^2\chi^2/M_{\mathcal{R}}^2\), whose presence is generic from the effective field theory point of view, and where \(M_{\mathcal{R}}\) denotes the energy scale associated with new physics above the energy scale of inflation \(H\).[^1] Along the inflationary valley, what can appear merely as a small correction to the kinetic term of \(\phi\) provides a negative, time-dependent contribution to the effective mass of \(\chi\), which reads \[\begin{aligned} m_{\mathrm{eff}}^2= m_h^2-4 \epsilon(t) H^2(t) M_\usssPl^2/M_{\mathcal{R}}^2, \label{eq:meff} \end{aligned}\] where \(M_\usssPl\) is the reduced Planck mass and \(\epsilon\equiv \dot{\phi}^2/(2 H^2 M_\usssPl^2)\) is the first slow-roll parameter. If \(\epsilon(t) H^2(t)\) increases during inflation, \(m_{\mathrm{eff}}^2\) can turn from positive to negative, triggering an instability when \(\epsilon\) reaches the critical value \[\begin{aligned} \label{eq:epsilonc} \epsilon_{ \mathrm{c}} = \frac{1}{4} \left(\frac{m_h}{H_\mathrm{c}}\right)^2 \left(\frac{M_\mathcal{R}}{M_\usssPl}\right)^2\,. \end{aligned}\] If \(\epsilon_{ \mathrm{c}}<1\), this instability occurs before inflation ends by slow-roll violation, and can lead to a premature ending of inflation, see for a discussion on the fate of the instability. As argued there, the energy scale \(M_\mathcal{R}\) associated to the field-space curvature can a priori lie anywhere between the Hubble scale and the Planck scale, so that \(\epsilon_{ \mathrm{c}}\) can be orders of magnitude smaller than one, and the GD arises in the bulk of the would-be inflationary phase. With the assumption of the instability prematurely ending inflation, a powerful lower bound on \(M_\mathcal{R}\) also arises from the requirement to obtain the correct power spectrum amplitude \(\mathcal{P}_\zeta = 2.2\times 10^{-9}\) of curvature fluctuations at the pivot scale \(k_* = 0.05\,\mathrm{Mpc}^{-1}\). At leading order in the slow-roll approximation, it is given by \(\mathcal{P}_\zeta = H_*^2/(8 \pi^2 M_\usssPl^2 \epsilon_*)\), where \(H_*\) and \(\epsilon_*\) are evaluated at the time when \(k_*\) crosses the Hubble radius during inflation (all the quantities with a subscript "\*" are evaluated at that time). At that time, \(m_{\mathrm{eff}}^2\) is still positive, which implies that \[\begin{aligned} \frac{M_\mathcal{R}}{H_*} > \frac{1}{\sqrt{2 \pi^2 \mathcal{P}_\zeta}} \frac{H_*}{m_h} \simeq 5000 \frac{H_*}{m_h}. \label{eq:lowerbound:MR} \end{aligned}\] Note that a similar bound on the mass scale characterizing the field space curvature is derived in beyond the simple setup described by the Lagrangian , which shows how the GD offers the possibility to constrain interactions at energies well above \(H\), and the internal geometry of high-energy physics theories. Notice that in the minimal realization used here, \(\epsilon(t) H^2(t)\), or equivalently \(\dot{\phi}^2\), must increase as time proceeds, which may not be satisfied for some inflationary potentials, e.g. \(V(\phi)\propto\phi^p\) with \(p\geq2\). However, in more general setups, \(m_h\) may receive corrections of order \(H\) (see e.g. ) or \(M_\mathcal{R}\) may depend on \(\phi\) and the GD may still provide a viable way to end inflation prematurely. To leave these possibilities open, we will consider a PEI for any choice of potential \(V(\phi)\), parameterize it by the value \(\epsilon_{\mathrm{c}}\) of the slow-roll parameter at the end of inflation, and use to interpret it in the context of the GD for the models that allow it.
## Reheating {#sec:reheating}
As mentioned in , the expansion history realized during reheating determines the amount of expansion realized between the Hubble exit time of the scales probed in the CMB and the end of inflation, hence the location of the observational window along the inflationary potential. More precisely, the number of \(e\)-folds \(\Delta N_*=\ln(a_\mathrm{end}/a_*)\) elapsed between Hubble exit time of the pivot scale \(k_*\) and the end of inflation is given by \[\begin{aligned} \Delta N_* = \frac{1-3\bar{w}_\mathrm{reh}}{12\left(1+\bar{w}_\mathrm{reh}\right)}\ln\left(\frac{\rho_\mathrm{reh}}{\rho_\mathrm{end}}\right) +\frac{1}{4}\ln\left(\frac{\rho_*}{9M_\usssPl^4}\frac{\rho_*}{\rho_\mathrm{end}}\right)-\ln\left(\frac{k_{*}/a_\mathrm{now}}{\tilde{\rho}_{\gamma,\,\mathrm{now}}^{1/4}}\right)\,. \label{eq:DeltaNstar} \end{aligned}\] In this expression, \(\bar{w}_\mathrm{reh}=\int_\mathrm{reh} w(N)\mathrm{d} N/N_\mathrm{reh}\) is the averaged equation-of-state parameter during reheating, \(\rho_\mathrm{reh}\) is the energy density of the Universe at the end of reheating, i.e. at the beginning of the radiation-dominated era, \(\rho_*\) is the energy density calculated \(\Delta N_*\) \(e\)-folds before the end of inflation, \(\rho_\mathrm{end}\) is the one at the end of inflation, \(a_\mathrm{now}\) is the present value of the scale factor, and \(\tilde{\rho}_{\gamma,\,\mathrm{now}}\) is the energy density of radiation today rescaled by the number of relativistic degrees of freedom. Reheating should occur after inflation and before big-bang nucleosynthesis (BBN), such that \(\rho_\mathrm{BBN}<\rho_\mathrm{reh}<\rho_\mathrm{end}\). Although in certain situations the reheating temperature may be below \(1\,\mathrm{MeV}\), in this work we make the conservative choice \(\rho^{1/4}_{\mathrm{BBN}} = 10\, \mathrm{MeV}\). Moreover, from the energy positivity conditions in general relativity, and the definition of reheating as the era following the accelerated phase of expansion, one has \(-1/3<\bar{w}_\mathrm{reh}<1\). By letting \(\rho_\mathrm{reh}\) and \(\bar{w}_\mathrm{reh}\) vary in their respective ranges, from one obtains an interval of possible values for \(\Delta N_*\), hence a certain uncertainty range along the inflationary potential. Notice that this interval depends on the inflationary potential, since in , the way \(\rho_*/\rho_\mathrm{end}\) is related to \(\Delta N_*\) depends on the inflationary history, hence on the potential, and the absolute value of \(\rho_*\) is constrained by \(\mathcal{P}_\zeta\), which also involves \(\epsilon_*\), that itself depends on the potential as well. To illustrate this effect and its degeneracy with a premature end of inflation, in , we have represented the relevant field values as a function of \(\epsilon_\mathrm{c}\) for two of the models studied in this work. For a given value of \(\epsilon_\mathrm{c}\), which determines the value of \(\phi_\mathrm{end}\), we can infer from the range of allowed values for \(\phi_*\) corresponding to the pivot scale \(k_*\) of the CMB; this range is displayed with the colored stripes. The location of that range depends both on the PEI parameter \(\epsilon_\mathrm{c}\) and on the reheating parameters through \(\Delta N_*\) given in . Naturally, the more premature the end of inflation is, the less important the effects of the uncertainties on the reheating expansion history are. The range \([\phi (\epsilon=1), \phi (\epsilon=\epsilon_\mathrm{c})]\) (below the line) is entirely removed from the inflationary dynamics, while the range \([\phi(\epsilon=\epsilon_\mathrm{c}), \phi_{*}^{\mathrm{min}}]\) (between the line and the stripe) is dynamically accessible but corresponds to smaller scales than the pivot scale. The range \([\phi_{*}^{\mathrm{min}}, \phi_{*}^{\mathrm{max}}]\) is probed at the pivot scale, where the uncertainty comes from our incomplete knowledge of reheating.
## Bayesian model comparison {#sec:Bayesian}
In order to discuss the impact of a PEI on identifying the favored models of inflation, we use the Bayesian inference techniques developed in . In this framework, for each model \(\mathcal{M}_i\) (labeled by \(i\)), the posterior probability \(p\) of its parameters \(\theta_{ij}\) (labeled by \(j\)) is expressed as \[\begin{aligned} \label{eq:posterior:def} p\left(\theta_{ij}\vert\mathcal{D},\mathcal{M}_i\right)=\frac{\mathcal{L} \left(\mathcal{D}\vert\theta_{ij},\mathcal{M}_i\right) \pi\left(\theta_{ij}\vert \mathcal{M}_i\right)}{\mathcal{E}\left(\mathcal{D}\vert\mathcal{M}_i \right) } \,. \end{aligned}\] In this expression, \(\mathcal{L}(\mathcal{D}\vert\theta_{ij},\mathcal{M}_i)\) is the likelihood and represents the probability of observing the data \(\mathcal{D}\) assuming the model \(\mathcal{M}_i\) is true and \(\theta_{ij}\) are the actual values of its parameters, \(\pi (\theta_{ij}\vert \mathcal{M}_i )\) is the prior distribution on the parameters \(\theta_{ij}\), and \(\mathcal{E}\left(\mathcal{D}\vert\mathcal{M}_i \right)\) is a normalization constant called the Bayesian evidence and defined as \[\begin{aligned} \label{eq:evidence:def} \mathcal{E}\left(\mathcal{D}\vert\mathcal{M}_i \right) = \int\mathrm{d}\theta_{ij}\mathcal{L} \left(\mathcal{D}\vert\theta_{ij},\mathcal{M}_i\right) \pi\left(\theta_{ij}\vert \mathcal{M}_i\right)\,. \end{aligned}\] The Bayesian evidence allows one to calculate the posterior probability of a model itself, \(p(\mathcal{M}_i\vert\mathcal{D})\propto \mathcal{E}(\mathcal{D}\vert\mathcal{M}_i)\pi(\mathcal{M}_i)\), where \(\pi(\mathcal{M}_i)\) is the prior assigned to the model. The posterior odds between two models \(\mathcal{M}_i\) and \(\mathcal{M}_j\) can then be written as \[\begin{aligned} \frac{p\left(\mathcal{M}_i\vert\mathcal{D}\right)} {p\left(\mathcal{M}_j\vert\mathcal{D}\right)} =\frac{\mathcal{E}\left(\mathcal{D}\vert\mathcal{M}_i\right)} {\mathcal{E}\left(\mathcal{D}\vert\mathcal{M}_j\right)} \frac{\pi\left(\mathcal{M}_i\right)}{\pi\left(\mathcal{M}_j\right)}\equiv B_{ij}\frac{\pi\left(\mathcal{M}_i\right)}{\pi\left(\mathcal{M}_j\right)}\,, \end{aligned}\] where we have defined the Bayes factor \(B_{ij}\) by \(B_{ij}=\mathcal{E}\left(\mathcal{D}\vert\mathcal{M}_i\right) /\mathcal{E}\left(\mathcal{D}\vert\mathcal{M}_j\right)\). Under the principle of indifference, one can assume non-committal model priors, \(\pi(\mathcal{M}_i)=\pi\left(\mathcal{M}_j\right)\), in which case the Bayes factor becomes identical to the posterior odds. With this assumption, a Bayes factor larger (smaller) than one means a preference for the model \(\mathcal{M}_i\) over the model \(\mathcal{M}_j\) (a preference for \(\mathcal{M}_j\) over \(\mathcal{M}_i\)). In practice, the "Jeffreys' scale" gives an empirical prescription for translating the values of the Bayes factor into strengths of belief. When \(\ln(B_{ij})>5\), \(\mathcal{M}_j\) is said to be "strongly disfavored" with respect to \(\mathcal{M}_i\), "moderately disfavored" if \(2.5<\ln(B_{ij})<5\), "weakly disfavored" if \(1<\ln(B_{ij})<2.5\), and the situation is said to be "inconclusive" if \(\ln(B_{ij})<1\). Bayesian analysis allows us to identify the models that achieve the best compromise between quality of the fit and lack of fine tuning. In practice, the data \(\mathcal{D}\) used in this work is the Planck 2015 \(TT\) data combined with the high-\(\ell\) \(C_\ell^{TE}+C_\ell^{EE}\) likelihood and the low-\(\ell\) temperature plus polarization likelihood (PlanckTT,TE,EE+lowTEB in the notations of , see table 1 there), together with the BICEP2-Keck/Planck likelihood described in , and the effective likelihood via slow-roll reparameterization of is employed. The predictions of the models are computed making use of the publicly available `ASPIC` library , which has been extended to incorporate the GD. An important aspect of Bayesian analysis is the role played by priors. For the GD sector, assuming that the scale of new physics \(M_{\mathcal{R}}\) is sub-Planckian, but leaving its order of magnitude undetermined, we adopt a Jeffreys prior (i.e. logarithmically flat) \(-25<\log_{10}(M_{\mathcal{R}}/M_\usssPl)<0\). The bound ([\[eq:lowerbound:MR\]](#eq:lowerbound:MR){reference-type="ref" reference="eq:lowerbound:MR"}) seems to impose an additional hard prior condition, but the normalization of the power spectrum automatically takes care of it. In practice, when \(M_{\mathcal{R}}/M_\usssPl>2 H_\mathrm{c}/m_h\), yields a value for \(\epsilon_\mathrm{c}\) that is larger than one, which means that the GD does not occur and inflation ends by slow-roll violation in a standard way. One may be concerned that the lower bound \(10^{-25}M_\usssPl\) corresponds to energy scales small enough to be probed in particle physics experiments. In a cosmological context however, the Friedmann equation \(H^2=\rho/(3M_\usssPl^2)\) involves the Planck mass and \(\rho_{\mathrm{BBN}}^{1/4} = 10\, \mathrm{MeV}\) yields a lower bound on \(H_*\) (hence on \(M_{\mathcal{R}}\)) of the order \(H_{\mathrm{BBN}} \sim 10^{-42} M_\usssPl\), so \(10^{-25}M_\usssPl\) is in fact more than conservative in this sense. In our parameterization, \(m_h\) only appears divided by \(H_\mathrm{c}\) in the expression ([\[eq:epsilonc\]](#eq:epsilonc){reference-type="ref" reference="eq:epsilonc"}) of \(\epsilon_\mathrm{c}\). In the following, we use the value \(m_h/H_\mathrm{c}=10\) but one should note that assuming \(m_h/H_\mathrm{c}\) to take another fixed value, or allowing it to vary with a logarithmically flat prior, would not modify the shape of the effective prior on \(\epsilon_\mathrm{c}\) induced by that is logarithmically flat. It would only change its lower bound but that would not affect our main conclusions. Let us also stress that using an effective logarithmically flat prior on \(\epsilon_\mathrm{c}\) allows us to scan premature end of inflation in general, beyond the phenomenon of the GD. For the reheating sector, \(\rho_{\mathrm{reh}}\) and \(\bar{w}_\mathrm{reh}\) only appear in through the combination given in the first term of the right-hand side denoted \(\ln R_\mathrm{rad} = (1-3 \bar{w}_\mathrm{reh})/(12+12\bar{w}_\mathrm{reh}) \ln(\rho_\mathrm{reh}/\rho_\mathrm{end})\). As explained in , \(\rho_\mathrm{reh}\) can vary between \(\rho_{\mathrm{BBN}}\) and \(\rho_\mathrm{end}\), and \(\bar{w}_\mathrm{reh}\) can vary between \(-1/3\) and \(1\), so that \(\ln(\rho_{\mathrm{BBN}}/\rho_\mathrm{end})/4<\ln R_\mathrm{rad}<\ln(\rho_\mathrm{end}/\rho_{\mathrm{BBN}})/12\). Since the order of magnitude of \(R_\mathrm{rad}\) is unknown between these two bounds, they define a logarithmically flat prior on \(R_\mathrm{rad}\). Thus far we have specified all priors except for the parameters specifying the model of inflation one considers. We now turn to the presentation of the prototypical models used in our analysis and the treatment of their free parameters.
## Prototypical models {#sec:models}
Instead of scanning all the \(\sim 200\) single-field models reported in and implemented in the `ASPIC` library, our strategy is to identify a few classes of models that behave differently under premature termination of inflation, and to study one or a few prototypical examples in each class. In order to discuss the predictions of these models and how they compare with the data, before a proper Bayesian analysis is performed in , in we show the induced priors of these models on \(n_\usssS\) and \(r\), where \(n_\usssS\) is the spectral index of the curvature fluctuation power spectrum \(\mathcal{P}_\zeta\), and \(r\) is the tensor-to-scalar ratio, both calculated at the pivot scale \(k_*\). In the slow-roll approximation, these are related to the slow-roll parameters through \(n_\usssS=1-2\epsilon_1-\epsilon_2-2\epsilon_1^2-(2C+3)\epsilon_1 \epsilon_2-C \epsilon_2 \epsilon_3\) and \(r= 16\epsilon_1 + 16C \epsilon_1 \epsilon_2\), where the parameter \(C \simeq-0.73\) is a numerical constant. These expressions are valid at second order in slow roll, and the slow-roll parameters \(\epsilon_1=\epsilon\), \(\epsilon_2\) and \(\epsilon_3\) can be calculated from the inflationary potential according to \(\epsilon_1 = M_\usssPl^2 (V^\prime/V)^2/2\), \(\epsilon_2 = 2 M_\usssPl^2 [(V^\prime/V)^2-V^{\prime\prime}/V]\) and \(\epsilon_3= 2M_\usssPl^4 [V^{\prime\prime\prime} V^\prime/V^2-3 V^{\prime\prime} {V^\prime}^2/V^3+ 2 (V^\prime/V)^4] /\epsilon_2\), where these expressions must be evaluated at \(\phi=\phi_*\). In practice, the induced priors are calculated using a fiducial constant likelihood in the `ASPIC` pipeline (one can check that plugging \(\mathcal{L}=\mathrm{constant}\) in and ([\[eq:evidence:def\]](#eq:evidence:def){reference-type="ref" reference="eq:evidence:def"}) gives \(p=\pi\) so that the posteriors extracted in this way indeed correspond to the induced priors).
### Plateau models {#sec:model:plateau}
A first class of models is made of plateau potentials that provide a good fit to the data in the standard setup where inflation ends by slow-roll violation. A typical example is the Starobinsky potential (SI) \[\begin{aligned} V\left(\phi\right) = M^4\left(1-e^{-\sqrt{\frac{2}{3}}\frac{\phi}{M_\usssPl}}\right)^2\,, \end{aligned}\] where, hereafter, the overall mass scale \(M^4\) is set to reproduce the correct normalization of \(\mathcal{P}_\zeta\). This potential does not have any free parameter so there is no prior to specify in the inflationary sector. The predictions of this model are displayed in the \((n_\usssS,r)\) plane in the top left panel of . Without a PEI, they fall into the current data sweet spot, with values of \(n_\usssS\) that can sometimes be too small but this corresponds to somewhat extreme reheating equation of state parameters \(\bar{w}_\mathrm{reh}\). This can be checked on the prior induced on \(n_\usssS\) alone displayed in the left panel of the second row in and which clearly peaks around \(n_\usssS\simeq 0.97\). When the PEI is added, the predictions extend to larger values of \(n_\usssS\) and smaller values of \(r\) that are disfavored by the data. The prior on \(n_\usssS\) is bimodal, with a first peak at the standard predictions and a second one around \(n_\usssS\simeq 1\), corresponding to very small values of \(\epsilon_\mathrm{c}\). The relative weights of these two modes depend on the exact lower bound on \(\epsilon_\mathrm{c}\) but one can expect the PEI to decrease the Bayesian evidence of this model in general.
### Large-field models {#sec:model:largefield}
A second class of models is made of large-field potentials that are currently disfavored since they predict values of \(r\) that are too large. A typical example is the monomial potential \[\begin{aligned} V\left(\phi\right) = M^4 \left(\frac{\phi}{M_\usssPl}\right)^p\,, \end{aligned}\] where in this work we consider \(p=2\) (\(\mathrm{LFI}_2\)) and \(p=4\) (\(\mathrm{LFI}_4\)). Their predictions are displayed in the middle and right top panels of , where one can check that in the absence of a premature termination of inflation, the value for \(r\) is indeed too large. When a PEI is allowed, \(r\) is made smaller, but this is at the expense of making \(n_\usssS\) too large, so one can expect these models to remain disfavored even when PEI is added.
### Hilltop models with a non-vanishing mass at the top {#sec:models:hilltopmass}
A third class of models consists in hilltop potentials with \(V^{\prime\prime}(\phi=0)\neq 0\). The first example of this class we consider is natural inflation (\(\mathrm{NI}\)) \[\begin{aligned} \label{eq:pot:ni} V= M^4 \left[ 1+\cos\left(\frac{\phi}{f}\right)\right]\,. \end{aligned}\] From a theoretical perspective, the parameter \(f\) is naturally sub-Planckian, but the model provides a good fit to the data only for \(f \gtrsim M_\usssPl\), which has motivated various mechanisms proposed in the literature to enlarge the value of \(f\) (see for a review). Here we leave the order of magnitude of \(f/M_\usssPl\) unspecified, and we work with a logarithmically flat prior \(-2<\log_{10}(f/M_\usssPl)<2\). Another example is small-field inflation 2 (\(\mathrm{SFI}_2\)) \[\begin{aligned} \label{eq:pot:sfi2} V\left(\phi\right) = M^4 \left[ 1-\left(\frac{\phi}{\mu}\right)^2\right]\,, \end{aligned}\] which also has a free parameter \(\mu\). If one Taylor expands the potential ([\[eq:pot:ni\]](#eq:pot:ni){reference-type="ref" reference="eq:pot:ni"}) of NI and compares it with the one ([\[eq:pot:sfi2\]](#eq:pot:sfi2){reference-type="ref" reference="eq:pot:sfi2"}) of \(\mathrm{SFI}_2\), one can identify \(\mu=2f\), which is why we adopt a logarithmically flat prior \(-2+\log_{10}(2)<\log_{10}(\mu/M_\usssPl)<2+\log_{10}(2)\) to allow for a fair comparison between these two models. Their predictions are displayed in the left and middle panels of the third row in . In the absence of a PEI, both models are brought to a good agreement with the data when \(f/M_\usssPl\) or \(\mu/M_\usssPl\) is large. When \(f/M_\usssPl\) or \(\mu/M_\usssPl\) decreases, \(r\) decreases but so does \(n_\usssS\) that quickly takes values that are too low. This is because as one approaches the top of the hill, the derivative of the potential becomes very small and so does \(r\), but the curvature of the potential saturates to a finite non-vanishing value, which yields a deviation from \(n_\usssS\) to \(1\) that increases when \(f\) or \(\mu\) decreases. When a PEI is allowed however, the opposite behavior is observed, since smaller values of \(r\) correspond to larger values of \(n_\usssS\), that interpolate between the ones favored by the data and \(1\), which is excluded. One may therefore expect the Bayesian evidence of these models not to change dramatically by allowing a PEI.
### Hilltop models with a vanishing mass at the top {#sec:models:hilltop:massless}
The behavior of hilltop models is different if \(V^{\prime\prime}(\phi=0) = 0\) and this constitutes our fourth class of models. A first example is small-field inflation 4 (\(\mathrm{SFI}_4\)) \[\begin{aligned} \label{eq:pot:sfi4} V\left(\phi\right) = M^4 \left[ 1-\left(\frac{\phi}{\mu}\right)^4\right]\,, \end{aligned}\] where by consistency with \(\mathrm{SFI}_2\), we use the logarithmically flat prior \(-2+\log_{10}(2)<\log_{10}(\mu/M_\usssPl)<2+2\log_{10}(2)\). The predictions of \(\mathrm{SFI}_4\) are displayed in the right panel of the third row of . They are similar to the ones of \(\mathrm{SFI}_2\) except that when \(r\) decreases, \(n_\usssS\) remains not too far from the observational constraints. When the PEI is allowed, \(n_\usssS\) is shifted towards larger values and intersects the ones preferred by the data, even if it shows preference for slightly too large values. Another example is the Colemann-Weinberg potential (\(\mathrm{CWI}_\mathrm{f}\)) \[\begin{aligned} \label{eq:pot:cwi} V\left(\phi\right) = M^4 \left[1+\alpha \left(\frac{\phi}{Q}\right)^4\ln\left(\frac{\phi}{Q}\right)\right]\,, \end{aligned}\] where \(\alpha=4e\) is a fixed constant set for the potential to vanish at its minimum. In the original version of the scenario, \(Q\) is fixed by the GUT scale, \(Q \sim 10^{14}-10^{15}\, \mathrm{GeV}\). It is therefore natural to choose a flat prior on \(Q\) (we denote this version of the scenario by \(\mathrm{CWI}_\mathrm{f}\), other versions are also considered in ), \(5\times 10^{-5} < Q/M_\usssPl < 5\times 10^{-4}\). The predictions of \(\mathrm{CWI}_\mathrm{f}\) are shown in the bottom left panel of , where one can see that the values predicted for \(n_\usssS\) are always too small without a PEI. When a PEI is allowed, \(n_\usssS\) is shifted towards larger values as in \(\mathrm{SFI}_4\), while \(r\) takes smaller values. The one-sigma contours (blue dashed lines) reveal that the distribution is bimodal and is peaked both at the predictions obtained without PEI and at values of \(n_\usssS\) close to one, both peaks being observationally disfavored.
### Inflection point models {#sec:inflection}
The fifth and last class of models we consider is made of potentials with a flat inflection point, such as MSSM inflation (\(\mathrm{MSSMI}_\mathrm{o}\)) \[\begin{aligned} \label{eq:pot:mssmi} V(\phi) = M^4 \left[\left(\frac{\phi}{\phi_0}\right)^2-\frac{2}{3}\left(\frac{\phi}{\phi_0}\right)^6+\frac{1}{5}\left(\frac{\phi}{\phi_0}\right)^{10}\right]\,. \end{aligned}\] The free parameter \(\phi_0\) can be expressed as \(\phi_0^8 = M_\usssPl^6 m_\phi^2/(10 \lambda_6^2)\), where \(\lambda_6\) is a coupling constant that is taken to be of order one, while \(m_\phi\) is a soft supersymmetry breaking mass and, thus, is chosen to be around \(\simeq 1\, \mathrm{TeV}\). One then obtains \(\phi_0\simeq 10^{14}\, \mathrm{GeV}\) and in the original form of this scenario (denoted \(\mathrm{MSSMI}_\mathrm{o}\), other versions are also considered in ), it is therefore natural to take a flat prior \(2\times 10^{-5} < \phi_0/M_\usssPl < 2 \times 10^{-4}\). Another example is the renormalizable inflection point inflation (\(\mathrm{RIPI}_\mathrm{o}\)) potential \[\begin{aligned} \label{eq:pot:ripi} V(\phi) = M^4 \left[\left(\frac{\phi}{\phi_0}\right)^2-\frac{4}{3}\left(\frac{\phi}{\phi_0}\right)^3+\frac{1}{2}\left(\frac{\phi}{\phi_0}\right)^4\right]\,, \end{aligned}\] with \(\phi_0=\sqrt{3}m_\phi/h\), where \(h\simeq 10^{-12}\) is a dimensionless coupling constant and \(m_\phi\) is a soft breaking mass of order \(100\, \mathrm{GeV}-10\, \mathrm{TeV}\). One then has \(\phi_0\simeq 10^{14}\, \mathrm{GeV}\) as in \(\mathrm{MSSMI}_\mathrm{o}\), and the same flat prior \(2\times 10^{-5} < \phi_0/M_\usssPl < 2 \times 10^{-4}\) can be used. The predictions of these models are displayed in the middle and right bottom panels of , where the situation is in fact similar to \(\mathrm{CWI}_\mathrm{f}\) but even more drastic since the version of the models without PEI predicts values of \(n_\usssS\) that are even more disfavored. Otherwise, the same remarks apply here.\
Before including the observational data in the Bayesian analysis, a final remark is in order. One may be concerned that allowing a PEI brings regions of the potential into the observational window that are so flat that they may be dominated by quantum diffusion effects , which would question the consistency of our classical slow-roll approach. In our treatment of the inflection point models presented in for instance, even when \(\epsilon_\mathrm{c}\) takes arbitrarily small values, the observational window is always located below the inflection point, since classically, it takes an infinite number of \(e\)-folds to cross the inflection point. However, quantum diffusion allows the field to cross the inflection point in a finite amount of time, so that one may wonder whether the PEI should extend the observational window to regions located above the inflection point when these stochastic corrections are taken into account. Stochastic effects dominate the field dynamics when the mean quantum kick over one \(e\)-fold, \(H/(2\pi)\), exceeds the classical drift \(V^{\prime}/(3H^2)\). Making use of the formula given above for \(\mathcal{P}_\zeta\), and recalling that \(\epsilon\simeq M_\usssPl^2 (V^\prime/V)^2/2\) in single-field slow-roll inflation, one can see that this happens when \(\mathcal{P}_\zeta>1\). Since the mass scale \(M^4\) in the above potentials is set precisely to satisfy the power spectrum normalization condition \(\mathcal{P}_\zeta(k_*)\simeq 2.2\times 10^{-9} \ll 1\), and since the models introduced above are all such that \(\mathcal{P}_\zeta\) decreases as inflation proceeds (they all feature red spectral indices \(n_\usssS<1\)), one is guaranteed that \(\mathcal{P}_\zeta \ll1\) between \(\phi_*\) and \(\phi_\mathrm{end}\), where the field therefore behaves classically, rendering our analysis consistent. In the case of inflection point models, this does not preclude other (larger) viable values of \(M^4\) from existing such that the power spectrum would be correctly normalized above the inflection point and that stochastic effects would allow the field to cross the inflection point in the correct finite amount of time, but these solutions are not included in our analysis.
# Results {#sec:results}
Let us now include the models introduced in in the Bayesian pipeline of `ASPIC`. In , their Bayesian evidence is given, together with the maximal value of the likelihood over their prior space, \(\mathrm{max}_{\theta_{ij}}\mathcal{L} \left(\mathcal{D}\vert\theta_{ij},\mathcal{M}_i\right)\), i.e. the "best fit" values. By definition, the best fit can only increase when a PEI is allowed since the parameter space extends. The Bayesian evidences have been normalized with respect to \(\mathrm{SI}\) without PEI for reference but the normalization choice is irrelevant. What matters is the change in the relative Bayesian evidences when allowing the PEI, which shows how PEI mechanisms can substantially reorder the ranking of inflationary models. The most striking change concerns inflection point models (\(\mathrm{MSSMI}_\mathrm{o}\) and \(\mathrm{RIPI}_\mathrm{o}\)) that are very strongly disfavored without a PEI but become almost weakly disfavored only when the PEI is allowed. We now analyze the different classes of models listed in in more details.
## Plateau models {#sec:result:plateau}
As expected in , the Bayesian evidence of the Starobinsky model (\(\mathrm{SI}\)), representative of plateau potentials that already provide a good fit to the data without a premature end of inflation, decreases when the PEI is allowed, by a amount \(\Delta \ln \mathcal{E}\simeq-2\). This is because the PEI explores regions of the potential that provide too large values for \(n_\usssS\). In the top left panel of , the posterior distribution on the mass scale \(M_\mathcal{R}\) associated with the curvature of the field space in the geometrical destabilization is displayed and translated into a posterior distribution on \(\epsilon_\mathrm{c}\). Since a PEI is disfavored, one obtains a lower bound on \(M_\mathcal{R}\) that reads \(\log_{10}(M_\mathcal{R}/M_\usssPl) >-2.94\) at the two-sigma confidence level, which translates into \(\epsilon_\mathrm{c}>3.3\times 10^{-5}\).
## Large-field models {#sec:result:largefield}
For large-field models that predict too large values of \(r\) in the standard setup, here represented by \(\mathrm{LFI}_2\) and \(\mathrm{LFI}_4\), the Bayesian evidence decreases when a PEI is allowed, and both models become strongly disfavored. As explained in , this is because, even though the PEI allows smaller values of \(r\) to be obtained, it is at the expense of larger values of \(n_\usssS\) that are even more disfavored by the data. However, as can be seen in the middle and top panels of , the prior in the \((n_\usssS,r)\) plane in these models comes closer to the observational contours with a PEI than without, with an improvement that is more pronounced for \(\mathrm{LFI}_4\) than for \(\mathrm{LFI}_2\), which is clearly visible in the best fits values of . This explains why, in the middle and top panels of , the posterior distribution on \(M_\mathcal{R}\) peaks at intermediate values, namely \(\log_{10}(M_\mathcal{R}/M_\usssPl)=-1.52\) for \(\mathrm{LFI}_2^\mathrm{PEI}\), corresponding to \(\epsilon_\mathrm{c} = 0.023\), and \(\log_{10}(M_\mathcal{R}/M_\usssPl)=-1.67\) for \(\mathrm{LFI}_4^\mathrm{PEI}\), corresponding to \(\epsilon_\mathrm{c} =0.012\), where the peak is even more pronounced for \(\mathrm{LFI}_4^\mathrm{PEI}\). However, because these peaks are very narrow, the values of \(M_\mathcal{R}\) leading to an improvement of the fit are fine tuned and this explains why the Bayesian evidences decrease.
## Hilltop models with a non-vanishing mass at the top {#sec:results:hilltopmass}
The Bayesian evidences of the hilltop models with a non-vanishing mass at the top of the hill is not strongly affected by the introduction of a PEI. \(\mathrm{NI}\) slightly improves and \(\mathrm{SFI}_2\) slightly worsens, but both models remain weakly disfavored with respect to \(\mathrm{SI}\). As noted in , this can be understood by examining the induced priors on \(n_\usssS\) and \(r\) displayed in the left and middle panels of the second row of , where one can see that for values of \(r\) larger than \(\sim 10^{-2}\), the two priors (with and without PEI) roughly coincide, at least at the one-sigma level, while for smaller values of \(r\), they scan disjoint, but both disfavored, parameter space regions (namely values of \(n_\usssS\) that are too large with the PEI and too low without). This is why a premature end of inflation does not change much the Bayesian status of these models. This is confirmed by studying the posterior distributions on \(M_\mathcal{R}\) in the left and middle panels of the second row in . In both cases, the distributions are rather flat, which is again consistent with the small impact PEI has on these models. For \(\mathrm{NI}^\mathrm{PEI}\), the distribution slightly peaks at an intermediate value of \(M_\mathcal{R}\), namely \(\ln_{10}(M_\mathcal{R}/M_\usssPl)=-2.79\), corresponding to \(\epsilon_\mathrm{c} = 6.7\times 10^{-5}\), and this peak is responsible for the slight increase in the Bayesian evidence. For \(\mathrm{SFI}_2^\mathrm{PEI}\) however, the posterior distribution is maximal around the standard value \(\epsilon_\mathrm{\mathrm{c}}\simeq 1\), which explains why the Bayesian evidence decreases when \(\epsilon_\mathrm{c}\) is allowed to vary. Another quantity of interest is the field value characterizing the width of the hill, namely \(f\) for \(\mathrm{NI}\) and \(\mu\) for \(\mathrm{SFI}_2\). The posterior distributions on these parameters is displayed in . In the absence of a PEI, one can check that only super-Planckian values of these parameters are allowed by the data, since values of the order or smaller than the Planck mass lead to values of \(n_\usssS\) that are too small. Generating a super-Planckian hill width in a consistent complete UV theory is not an easy task in these models and has been the subject of an abundant literature. One can see that allowing a PEI does not alleviate this problem since Planckian or sub-Planckian values of \(f\) and \(\mu\) are still strongly disfavored. However, it removes the super-Planckian tail of the distributions and leads to a clear measurement of these parameters, namely \(\log_{10}(f/M_\usssPl)=0.77\pm0.24\) for \(\mathrm{NI}^\mathrm{PEI}\) and \(\log_{10}(\mu/M_\usssPl)=1.08\pm 0.19\) for \(\mathrm{SFI}_2^\mathrm{PEI}\) at the one-sigma level.[^2]
## Hilltop models with a vanishing mass at the top {#hilltop-models-with-a-vanishing-mass-at-the-top}
Hilltop models with a vanishing mass at the top of the hill are substantially more affected by PEI mechanisms, although the effect depends on the details of the potential one considers. For \(\mathrm{SFI}_4\), the PEI decreases the Bayesian evidence of the model, which becomes weakly disfavored. This is because the spectral index \(n_\usssS\), which is slightly too small (but still compatible with the data) in the standard setup, is mostly predicted to be close to scale invariance when the PEI is allowed, as can be seen in the one-sigma contours of the right panel of the third row of , which is observationally excluded. This is why, contrary to \(\mathrm{SFI}_2\) discussed in , small values of \(M_\mathcal{R}\) are excluded as can be seen in the posterior distribution of \(M_\mathcal{R}\) displayed in the middle right panel of . More precisely, one gets the two-sigma constraint \(\log_{10}(M_\mathcal{R}/M_\usssPl)>-7.8\), corresponding to \(\epsilon_\mathrm{c}>4.5\times 10^{-6}\). As for \(\mathrm{SFI}_2\), it is also worth discussing the posterior distribution on \(\mu\), displayed in the top right panel of . In the absence of a PEI, and contrary to \(\mathrm{SFI}_2\), one can see that sub-Planckian values of \(\mu\) are marginally allowed since they give rise to values of \(n_\usssS\) that are not excluded by the data. However, the model still features some preference for Planckian or slightly super-Planckian values of \(\mu\) or order \(\mu\sim 10M_\usssPl\). When the PEI is allowed however, the sub-Planckian tail of the distribution is lifted up to a plateau which overall shows preference for sub-Planckian values of \(\mu\). This is in sharp contrast with \(\mathrm{SFI}_2\) and sheds new light on the problem of getting super-Planckian hill widths since in \(\mathrm{SFI}_4^\mathrm{PEI}\), this is not a requirement anymore. Let us however note that this is at the expense of making the model weakly disfavored overall. For \(\mathrm{CWI}_{\mathrm{f}}\), moderately disfavored in the standard case since it predicts values of \(n_\usssS\) that are too low, the model improves when a PEI is allowed and becomes weakly disfavored only. From the induced prior on \(n_\usssS\) and \(r\) displayed in the bottom left panel of , we already noted in that with a PEI, either \(n_\usssS\) is predicted to be at the level obtained in the standard setup, which is too low, or it is predicted to be close to scale invariance, which is too large. This is why a PEI does not fully succeed in making the model favored. In between these two peaks of the prior distribution, the predictions sweep the data's sweet spot and this is why intermediate values of \(M_\mathcal{R}\) are strongly preferred in the posterior distribution displayed in the bottom left panel of . It is interesting to notice that \(M_\mathcal{R}\) is accurately measured in this model, and one obtains \(\log_{10}(M_\mathcal{R}/M_\usssPl)=-11.1 \pm 2.9\) at the one-sigma confidence level, corresponding to \(\log_{10}(\epsilon_\mathrm{c}) =-20.7 \pm 7.2\). The posterior distribution on \(Q\), the parameter appearing in the potential ([\[eq:pot:cwi\]](#eq:pot:cwi){reference-type="ref" reference="eq:pot:cwi"}), is displayed in the bottom left panel of but is weakly constrained with or without PEI.
## Inflection point models {#sec:results:inflection}
The situation of inflection point models is very similar to the one of \(\mathrm{CWI}_\mathrm{f}\) but even more pronounced, since the values predicted for \(n_\usssS\) are even smaller without PEI than the ones for \(\mathrm{CWI}_\mathrm{f}\). This explains why the shift in the Bayesian evidence of \(\mathrm{MSSMI}_\mathrm{o}\) and \(\mathrm{RIPI}_\mathrm{o}\) is even larger than the one observed for \(\mathrm{CWI}_\mathrm{f}\), even though the Bayesian evidences of these three models with PEI is comparable. As for \(\mathrm{CWI}_\mathrm{f}\), the posterior distribution of \(M_\mathcal{R}\) for \(\mathrm{MSSMI}_\mathrm{o}\) and \(\mathrm{RIPI}_\mathrm{o}\), displayed in the bottom middle and right panels of , has a sharp peak, which leads to a measurement of \(M_\mathcal{R}\). One finds \(\log_{10}(M_\mathcal{R}/M_\usssPl) =-18.3\pm 0.8\) for \(\mathrm{MSSMI}_\mathrm{o}\), corresponding to \(\log_{10}(\epsilon_\mathrm{c}) =-35.2\pm 3.1\), and \(\log_{10}(M_\mathcal{R}/M_\usssPl) = -17.6 \pm 0.8\) for \(\mathrm{RIPI}_\mathrm{o}\), corresponding to \(\log_{10}(\epsilon_\mathrm{c}) =-33.8\pm 3.1\). The posterior distributions on \(\phi_0\), the parameter appearing in the potentials ([\[eq:pot:mssmi\]](#eq:pot:mssmi){reference-type="ref" reference="eq:pot:mssmi"}) and ([\[eq:pot:ripi\]](#eq:pot:ripi){reference-type="ref" reference="eq:pot:ripi"}), are displayed in the bottom right panels of but are weakly constrained with or without GD. Eventually, it is interesting to translate the constraints on \(M_\mathcal{R}\) on the derived parameter[^3] \(M_\mathcal{R}/H_*\), obtaining for both models \(\log_{10}(M_\mathcal{R}/H_*) = 3.0\pm 0.2\). A few comments are now in order. We studied two inflection point models only, but they are typical for models of inflation arising in string theory. With \(H_*\sim\mathcal{O}(\mathrm{MeV})\), these are models of low-scale inflation. Values of \(M_\mathcal{R}\sim\mathcal{O}(\mathrm{GeV})\) may be considered extremely low for a cutoff scale in the effective dimension-6 operator \((\partial\phi)^2\chi^2/M_\mathcal{R}^2\) from a particle physics point of view, but this is just another incarnation of extreme fine-tuning present in these models. In fact, this scale of high-energy effects lies three orders of magnitude above the scale of inflation \(H_*\) and it is quite remarkable that it can be constrained observationally without resorting to primordial non-Gaussianities.
# Discussion and conclusion {#sec:concl}
Motivated by the mechanism of the geometrical destabilization of inflation, we have investigated in this work how the Bayesian ranking of single-field slow-roll models of inflation is affected when allowing a mechanism of premature ending of inflation. We have found that plateau potentials that already provide a good fit to the data can only be made worse in the presence of a premature termination of inflation, and that large-field models that lead to values of the tensor-to-scalar ratio \(r\) that are too large in the standard setup are still disfavored when a PEI is allowed since the reduction of \(r\) that it yields is to the detriment of a too large increase in the value of the scalar spectral index \(n_\usssS\). Quadratic hilltop models, that predict values of \(n_\usssS\) that are too low when the hill has a sub-Planckian width, are not largely affected by the PEI. This is because, even though the PEI increases the values of \(n_\usssS\), it allows the models to match observational constraints only in a fine-tuned range of the parameter space, and otherwise yields values of \(n_\usssS\) that are too large. Quartic hilltop models on the other hand can be more substantially affected by a PEI, in a way that however depends on the details of the potential. In the case of \(\mathrm{SFI}_4\) where \(V\propto 1-(\phi/\mu)^4\), contrary to the standard case, a PEI favors sub-Planckian values of \(\mu\), that are more natural in these models. Finally, inflection point models which predict values of \(n_\usssS\) that are too low in the standard case and are therefore strongly disfavored, are only weakly disfavored when a PEI is allowed. In this case, and when interpreted in the framework of the geometrical destabilization, sharp measurements of the field-space curvature mass scale \(M_\mathcal{R}\), at the level of the \(\mathrm{GeV}\) scale, were derived. These results demonstrate how the interpretation of cosmological data in terms of fundamental physics and model building can be drastically modified in the presence of a premature end of inflation, as motivated by the mechanism of the geometrical destabilization. By discussing a few classes of models, more involved behaviors can also be addressed by viewing our prototypical examples as building blocks for more complicated phenomenologies. For example, \(\alpha\)-attractor models , which interpolate between the Starobinsky and large-field potentials, can be discussed by combining the results of and [3.2](#sec:result:largefield){reference-type="ref" reference="sec:result:largefield"}. Since both the Starobinsky model and the large-field potentials become worse in the presence of a PEI, \(\alpha\)-attractors are most certainly also worsened by allowing a PEI. Let us also note that the situation of models predicting a value for \(n_\usssS\) that is too large in the standard case has not been explicitly discussed so far. In fact, two cases can be distinguished, depending on whether the value predicted for \(n_\usssS\) is too large but still red, \(n_\usssS<1\), or blue, \(n_\usssS>1\). Not many models fall in the first category, the typical example being power-law inflation for which \(V(\phi)\propto \exp(-\alpha\phi/M_\usssPl)\). This potential is however conformally invariant so that changing the end of inflation location has exactly no impact on the predictions of the model. In the second case, \(n_\usssS>1\), \(\epsilon\) decreases as inflation proceeds (this is because in single-field slow-roll inflation, \(n_\usssS\simeq 1-2 \epsilon-\mathrm{d}\ln\epsilon/\mathrm{d} N\) and \(\epsilon\) is always positive), which means that \(m_{\mathrm{eff}}^2\) in increases and the GD cannot take place. So a PEI has to be realized through another mechanism. In that case, a potential of the form \(V\propto 1+(\phi/\phi_0)^p\), which predicts \(n_\usssS>1\) at \(\phi\ll \phi_0\) and \(n_\usssS<1\) at \(\phi\gg \phi_0\), could be turned from blue to red if inflation ends prematurely. However, in the \(\phi\gg \phi_0\) regime, the model asymptotes large-field inflation, which has been shown to be disfavored with or without PEI in . Since the model interpolates between two disfavored limits, it is likely disfavored as well. An important aspect of a PEI is that the shift in the observational window it induces is degenerate with uncertainties about reheating, which determines the location of the observational window with respect to the end of inflation. This is why, as explained in , it is important to properly account for the role played by reheating in the analysis. Conversely, introducing a premature end of inflation mechanism also leads to different constraints on the reheating epoch itself. In , we have shown the posterior distributions on the number of \(e\)-folds \(\Delta N_*\) elapsed between Hubble exit of the CMB pivot scale and the end of inflation for the models studied in this work, with and without PEI. One can see that in general, PEI allows for a wider range of values of \(\Delta N_*\) to be realized, and shows preference for smaller values than the ones obtained in the standard setup where inflation ends by slow-roll violation. This is mainly due to the two following reasons. First, small values of \(\Delta N_*\) that are disfavored in the standard setup since they correspond to parts of the inflationary potential too close to the end of inflation where \(\epsilon\sim 1\), hence too steep, can be allowed when a PEI is introduced since \(\epsilon_\mathrm{c}\) can be much smaller than one then. Second, in , one can see that \(\Delta N_*\) depends on the absolute energy scale of inflation through \(\ln(\rho_*/M_\usssPl^4) = \ln(24 \pi^2 \mathcal{P}_\zeta \epsilon_*)\) in the second term of the right-hand side, where we have used the expression given above for the power spectrum of scalar perturbations \(\mathcal{P}_\zeta\) together with Friedmann equation. Since \(\epsilon_*\) is smaller than \(\epsilon_\mathrm{c}\) for the models considered in this work, \(\epsilon_*\) is typically much smaller when a premature end of inflation is allowed, which also explains why \(\Delta N_*\) is smaller. The results derived in this work can be interpreted at two different levels. At the first level, the parameterization adopted to describe the GD induces an effective logarithmically flat prior on the first slow-roll parameter at the end of inflation \(\epsilon_\mathrm{c}\) as explained in , so that phenomenological and generic constraints about the end of inflation were derived, beyond the mechanism of the GD. They revealed that the Bayesian status of inflationary models can be substantially affected in the presence of a premature termination of inflation, for instance in inflection point models where sharp observational constraints on \(\epsilon_\mathrm{c}\) were derived. At the second level, within the framework of the GD, \(\epsilon_\mathrm{c}\) is related to the mass of the auxiliary field and to the field-space curvature along the inflationary valley. The constraints obtained on \(\epsilon_\mathrm{c}\) can therefore be translated into constraints or measurements on these parameters. Interestingly, we found that these constraints can be quite sharp. For the inflection point models \(\mathrm{MSSMI}_\mathrm{o}\) and \(\mathrm{RIPI}_\mathrm{o}\) for instance, with \(m_h=10 H_\mathrm{c}\), one obtains \(\log_{10}(M_\mathcal{R}/H_*) = 3.0\pm 0.2\). In other words, one can observationally constrain, and even pinpoint, high-energy effects that lie orders of magnitude above the energy scale of inflation, without resorting to primordial non-Gaussianities. This shows how the investigation of ultraviolet effects in the inflationary dynamics, such as the geometrical destabilization where the field-space geometry plays an important role, is crucial to further extend the range of energy scales that are accessible through cosmological surveys. | {'timestamp': '2017-11-06T02:10:14', 'yymm': '1706', 'arxiv_id': '1706.01835', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01835'} |
# The Higher Order \(G\)-Transformation {#se1}
The \(G\)-transformation was designed by Gray and Atchison as an extrapolation method for evaluating infinite integrals of the form \(\int^{\infty}_a f(t)\,dt\equiv I[f]\). It was later generalized in different ways in Atchison and Gray and Gray and Atchison, the ultimate generalization being given in Gray, Atchison, and McWilliams. This generalization was denoted the higher-order \(G\)-transformation. The way it is defined in, this transformation produces approximations to \(I[f]\) that are of the form \[\label{eq:1} G_n(x;h)= \frac{\left| \begin{array}{cccc} F(x)&F(x+h)&\cdots&F(x+nh)\\ f(x)&f(x+h)&\cdots&f(x+nh)\\ \vdots&\vdots&&\vdots \\ f(x+(n-1)h)&f(x+nh)&\cdots&f(x+(2n-1)h) \end{array} \right|}{\left|\begin{array}{cccc} 1&1&\cdots&1\\ f(x)&f(x+h)&\cdots&f(x+nh)\\ \vdots&\vdots&&\vdots \\ f(x+(n-1)h)&f(x+nh)&\cdots&f(x+(2n-1)h) \end{array} \right|},\] where \[\label{eq:2} F(x)=\int^x_af(t)\, dt.\] The approximations produced by the \(G\)-transformation of are simply the \(G_1(x;h)\). It follows from ([\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"}) that \(G_n(x;h)\) is also the solution of the linear system \[\label{eq:3} F(x+ih)=G_n(x;h)+\sum^n_{k=1}\bar{\alpha}_k f(x+(i+k-1)h), \ \ i=0,1,\ldots, n,\] where \(\bar{\alpha}_k\) are additional unknowns. It has been shown in that the kernel of the higher-order \(G\)-transformation is the set of functions \(f(x)\) that are integrable at infinity in the sense of Abel and that satisfy linear homogeneous ordinary differential equations of order \(n\) with constant coefficients. Thus \(f(x)\) is in this kernel if it is of the form \(f(x)=\sum^r_{k=1}p_k(x)e^{c_k x}\), where the \(c_k\neq 0\) are distinct and \(\Re c_k\leq 0\), \(p_k(x)\) are polynomials. If \(\mu_k\) is the degree of \(p_k(x)\) for each \(k\), and if \(\sum^r_{k=1}(\mu_k+1)=n\), then \(G_n(x;h)=I[f]\) for all \(x\) and \(h\). On the basis of this result it was concluded in Levin and Sidi that the higher-order \(G\)-transformation will be effective on functions of the form \(f(x)=\sum^s_{k=1}e^{c_kx}v_k(x)\), where \(v_k(x)\sim \sum^{\infty}_{i=0}\alpha_{ki}x^{\gamma_k-i}\) as \(x\to\infty\), with arbitrary \(\gamma_k\), provided \(h\) is of a suitable size. In the present work, we are concerned with the actual computation of the \(G_n(x;h)\). Of course, it is not desirable to compute \(G_n(x;h)\) via the determinantal representation in ([\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"}). Direct solution of the linear systems in ([\[eq:3\]](#eq:3){reference-type="ref" reference="eq:3"}) is expensive too. A very efficient and elegant procedure for computing the \(G_n(x;h)\) was given by Pye and Atchison in, and it has been denoted the rs-algorithm in Brezinski and Redivo Zaglia. The derivation of the rs-algorithm makes use of the representation in ([\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"}). In the present work, we develop yet another procedure for computing the \(G_n(x;h)\), and we call this new procedure the FS/qd-algorithm. We show that the FS/qd-algorithm is more efficient than the rs-algorithm. With proper substitutions to be discussed in Section [3](#se3){reference-type="ref" reference="se3"}, the FS/qd-algorithm can also be used to implement the transformation of Shanks, which is normally implemented via the well-known \(\varepsilon\)-algorithm of Wynn. We show that, when used for implementing the transformation of Shanks, the FS/qd-algorithm compares very favorably with the \(\varepsilon\)-algorithm.
# Algorithms for the Higher Order \(G\)-Transformation {#se2}
We start with a review of the algorithm of Pye and Atchison. Actually these authors consider the more general problem in which one would like to compute the quantities \(A^{(j)}_n\) defined via the linear equations \[\label{eq:4} A_l=A^{(j)}_n+\sum^n_{k=1}\bar{\alpha}_ku_{k+l-1},\ \ l=j,j+1,\ldots, j+n,\] where the \(A_i\) and \(u_i\) are known scalars, while the \(\bar{\alpha}_k\) are not necessarily known. Comparing ([\[eq:3\]](#eq:3){reference-type="ref" reference="eq:3"}) with ([\[eq:4\]](#eq:4){reference-type="ref" reference="eq:4"}) we can draw the analogy \(A_i\leftrightarrow F(x+ih)\), \(u_i\leftrightarrow f(x+ih)\), and \(A^{(j)}_n\leftrightarrow G_n(x+jh;h)\). The algorithm of computes the \(A^{(j)}_n\) with the help of two sets of auxiliary quantities, \(r^{(j)}_n\) and \(s^{(j)}_n\). These quantities are defined by \[\label{eq:5} r^{(j)}_n=\frac{H^{(j)}_n}{K^{(j)}_n},\ \ s^{(j)}_n=\frac{K^{(j)}_{n+1}}{H^{(j)}_n},\] where \(H^{(j)}_n\), the Hankel determinant associated with \(\{u_s\}\), and \(K^{(j)}_n\) are given as in \[\label{eq:6} H^{(j)}_n=\left|\begin{array}{cccc} u_{j}&u_{j+1}&\cdots&u_{j+n-1}\\ u_{j+1}&u_{j+2}&\cdots&u_{j+n}\\ \vdots&\vdots &&\vdots\\ u_{j+n-1}&u_{j+n}&\cdots&u_{j+2n-2} \end{array}\right|,\ \ H^{(j)}_0=1,\] and \[\label{eq:7} K^{(j)}_n=\left|\begin{array}{cccc} 1&1&\cdots&1\\ u_{j}&u_{j+1}&\cdots&u_{j+n-1}\\ u_{j+1}&u_{j+2}&\cdots&u_{j+n}\\ \vdots&\vdots &&\vdots\\ u_{j+n-2}&u_{j+n-1}&\cdots&u_{j+2n-3} \end{array}\right|, \ \ K^{(j)}_0=1.\] The rs-algorithm computes the \(r^{(j)}_n\), and \(s^{(j)}_n\) simultaneously by efficient recursions. Once these have been computed the \(A^{(j)}_n\) can be computed via a separate recursion.\
1. Set \[s^{(j)}_0=1,\ \ r^{(j)}_1=u_j,\ \ A^{(j)}_0=A_j,\ \ j=0,1,\ldots\.\]
2. For \(j=0,1,\ldots,\) and \(n=1,2,\ldots,\) compute recursively \[s^{(j)}_{n}=s^{(j+1)}_{n-1}\left(\frac{r^{(j+1)}_{n}}{r^{(j)}_{n}}-1 \right), \ \ r^{(j)}_{n+1}=r^{(j+1)}_n\left(\frac{s^{(j+1)}_{n}} {s^{(j)}_{n}}-1\right).\]
3. For \(j=0,1,\ldots,\) and \(n=1,2,\ldots,\) set \[A^{(j)}_n=\frac{r^{(j)}_{n}A^{(j+1)}_{n-1}-r^{(j+1)}_nA^{(j)}_{n-1}} {r^{(j)}_{n}-r^{(j+1)}_n}.\]
Before proceeding further, we note that the equations in ([\[eq:4\]](#eq:4){reference-type="ref" reference="eq:4"}) are the same as \[\label{eq:8} A_l=A^{(j)}_n+\sum^n_{k=1}\bar{\alpha}_kg_k(l),\ \ l=j,j+1,\ldots, j+n,\] with \(g_k(l)=u_{k+l-1}\), \(l=0,1,\ldots\.\) Linear systems as in ([\[eq:8\]](#eq:8){reference-type="ref" reference="eq:8"}) arise also in the definition of a generalized Richardson extrapolation method. This suggests that the E-algorithm of Schneider and the FS-algorithm of Ford and Sidi can be used for computing the \(A^{(j)}_n\). Different derivations of the E-algorithm were given by Håvie and Brezinski. Of course, direct application of these algorithms without taking into account the special nature of the \(g_k(l)\) is very uneconomical. By taking the nature of the \(g_k(l)\) into consideration, it becomes possible to derive fast algorithms for the \(A^{(j)}_n\). Now the E-algorithm produces the \(A^{(j)}_n\) by a recursion relation of the form \[A^{(j)}_n=\frac{R^{(j)}_{n}A^{(j+1)}_{n-1}-R^{(j+1)}_nA^{(j)}_{n-1}} {R^{(j)}_{n}-R^{(j+1)}_n},\] the most expensive part of the algorithm being the determination of the \(R^{(j)}_{n}\). Comparing the known expression for \(R^{(j)}_{n}\) when \(g_k(l)=u_{k+l-1}\) with that of \(r^{(j)}_{n}\), we realize that \(R^{(j)}_n=(-1)^{n-1}r^{(j)}_{n}\). We thus conclude that the rs-algorithm is simply the E-algorithm in which the \(R^{(j)}_{n}\), whose determination forms the most expensive part of the E-algorithm, are computed by a fast recursion. For this point and others, see. In view of the close connection between the rs-and E-algorithms, it is natural to investigate the possibility of designing another algorithm that is related to the FS-algorithm. This is worth the effort as the FS-algorithm is more economical than the E-algorithm to begin with. To this end we start with a brief description of the FS-algorithm and refer the reader to for details. For a comprehensive summary, see also Sidi. Let us first define the short-hand notation \[\label{eq:9} |u_1(j)\, u_2(j)\,\cdots\, u_n(j)|=\left |\begin{array}{cccc} u_1(j)& u_2(j)&\cdots&u_n(j) \\ u_1(j+1)& u_2(j+1)&\cdots&u_n(j+1)\\ \vdots&\vdots&&\vdots\\ u_1(j+n-1)&u_2(j+n-1)&\cdots&u_n(j+n-1) \end{array}\right|,\] and set \[\label{eq:10} G^{(j)}_n=|g_1(j)\, g_2(j)\,\cdots\, g_n(j)|, \ \ G^{(j)}_0=1.\] Next, let us agree to denote the sequence \(\{b(l)\}^{\infty}_{l=0}\) by \(b\) for short, and define \[\label{eq:11} f^{(j)}_n(b)=|g_1(j)\, g_2(j)\,\cdots\, g_n(j)\, b(j)|, \ \ f^{(j)}_0(b)=b(j).\] Finally, let us define \[\label{eq:12} \psi^{(j)}_n(b)=\frac{f^{(j)}_n(b)}{G^{(j)}_{n+1}}.\] Then we have \[\label{eq:13} A^{(j)}_n=\frac{\psi^{(j)}_n(a)}{\psi^{(j)}_n(I)},\] where \(a\) and \(I\) denote the sequences \(\{a(l)=A_l\}^{\infty}_{l=0}\) and \(\{I(l)=1\}^{\infty}_{l=0}\) respectively. The FS-algorithm computes the quantities \(\psi^{(j)}_n(b)\) by a recursion of the form \[\label{eq:14} \psi^{(j)}_n(b)=\frac{\psi^{(j+1)}_{n-1}(b)-\psi^{(j)}_{n-1}(b)} {D^{(j)}_n}, \ \ D^{(j)}_n=\frac{G^{(j)}_{n+1}G^{(j+1)}_{n-1}} {G^{(j)}_{n}G^{(j+1)}_{n}}.\] By this recursion, first the \(\psi^{(j)}_n(a)\) and \(\psi^{(j)}_n(I)\) are computed and then \(A^{(j)}_n\) is determined via ([\[eq:13\]](#eq:13){reference-type="ref" reference="eq:13"}). We recall that the most expensive part of the FS-algorithm is the (recursive) determination of the quantities \(D^{(j)}_n\), and we would like to reduce the cost of this part. Fortunately, this can be achieved once we realize that, with \(G^{(j)}_n\) and \(H^{(j)}_n\) defined as in ([\[eq:10\]](#eq:10){reference-type="ref" reference="eq:10"}) and ([\[eq:6\]](#eq:6){reference-type="ref" reference="eq:6"}) respectively, and with \(g_k(l)=u_{k+l-1}\), we have \(G^{(j)}_n=H^{(j)}_n\) in the present case. From this and ([\[eq:14\]](#eq:14){reference-type="ref" reference="eq:14"}), we obtain the surprising result that \[\label{eq:15} D^{(j)}_n=\frac{H^{(j)}_{n+1}H^{(j+1)}_{n-1}} {H^{(j)}_{n}H^{(j+1)}_{n}}=e^{(j)}_n.\] Here \(e^{(j)}_n\) is a quantity computed by the famous qd-algorithm of Rutishauser, a clear exposition of which can also be found in Henrici. Actually, the qd-algorithm computes along with the \(e^{(j)}_n\) also the quantities \(q^{(j)}_n\) that are given as in \[\label{eq:16} q^{(j)}_n=\frac{H^{(j)}_{n-1}H^{(j+1)}_{n}} {H^{(j)}_{n}H^{(j+1)}_{n-1}}.\] The \(q^{(j)}_n\) and \(e^{(j)}_n\) serve the construction of the regular \(C\)-fractions, hence the Padé approximants, associated with the formal power series \(\sum^{\infty}_{k=0}u_kz^k\). Regular \(C\)-fractions are continued fractions of a special type. (See Jones and Thron.) The qd-algorithm computes the \(q^{(j)}_n\) and \(e^{(j)}_n\) via the recursions \[\label{eq:17} e^{(j)}_n=q^{(j+1)}_n-q^{(j)}_n+e^{(j+1)}_{n-1},\ \ q^{(j)}_{n+1}=\frac{e^{(j+1)}_{n}}{e^{(j)}_n}q^{(j+1)}_{n}, \ \ j\geq 0,\ n\geq 1,\] with the initial conditions \(e^{(j)}_0=0\) and \(q^{(j)}_1=u_{j+1}/u_j\) for all \(j\geq 0\). The quantities \(q^{(k)}_n\) and \(e^{(k)}_n\) can be arranged in a two-dimensional array as in Figure [\[fig\]](#fig){reference-type="ref" reference="fig"}. This observation enables us to combine the FS-and qd-algorithms to obtain the following economical implementation, the FS/qd-algorithm, for the higher-order \(G\)-transformation. For simplicity of notation, we will let \(\psi^{(j)}_n(a)=M^{(j)}_n\) and \(\psi^{(j)}_n(I)=N^{(j)}_n\) in the FS-algorithm.\
1. Set \[e^{(j)}_0=0,\ \ q^{(j)}_1=\frac{u_{j+1}}{u_j},\ \ M^{(j)}_0=\frac{A_j}{u_{j}},\ \ N^{(j)}_0=\frac{1}{u_{j}}.\]
2. For \(j=0,1,\ldots,\) and \(n=1,2,\ldots,\) compute recursively \[e^{(j)}_n=q^{(j+1)}_n-q^{(j)}_n+e^{(j+1)}_{n-1},\ \ q^{(j)}_{n+1}=\frac{e^{(j+1)}_{n}}{e^{(j)}_n}q^{(j+1)}_{n},\] \[M^{(j)}_n=\frac{M^{(j+1)}_{n-1}-M^{(j)}_{n-1}} {e^{(j)}_n}, \ \ N^{(j)}_n=\frac{N^{(j+1)}_{n-1}-N^{(j)}_{n-1}} {e^{(j)}_n}.\]
3. For \(j=0,1,\ldots,\) and \(n=1,2,\ldots,\) set \[A^{(j)}_n=\frac{M^{(j)}_n}{N^{(j)}_n}.\]
It seems that, when a certain number of the \(u_i\) are given, it is best to compute the associated qd-table columnwise. (Note that the quantities in each of the recursions for \(e^{(j)}_n\) and \(q^{(j)}_{n+1}\) in step 2 of the FS/qd-algorithm above form the four corners of a lozenge in Figure [\[fig\]](#fig){reference-type="ref" reference="fig"}.) Following that we can compute the \(M^{(j)}_n\), \(N^{(j)}_n\), and \(A^{(j)}_n\) columnwise as well.
# Comparison of the rs-and FS/qd-Algorithms {#se3}
Let us now compare the operation counts of the two algorithms. First, we note that the \(r^{(j)}_n\) and \(s^{(j)}_n\) in the rs-algorithm can be arranged in a table similar to the qd-table of the \(e^{(j)}_n\) and \(q^{(j)}_n\). Thus, given \(A_0,A_1,\ldots,A_L\), and \(u_0,u_1,\ldots,u_{2L}\), we can compute \(A^{(j)}_n\) for \(0\leq j+n\leq L\). Now the number of the \(e^{(j)}_n\) in the relevant qd-table is \(L^2+O(L)\) and so is that of the \(q^{(j)}_n\). A similar statement can be made concerning the \(r^{(j)}_n\) and \(s^{(j)}_n\). The number of the \(A^{(j)}_n\) is \(L^2/2+O(L)\), and so are the numbers of the \(M^{(j)}_n\) and the \(N^{(j)}_n\). Consequently, we have the following operation counts.
Algorithm No. of Multiplications No. of Additions No. of Divisions
----------- ------------------------ ------------------ ------------------
FS/qd \(L^2+O(L)\) \(3L^2+O(L)\) \(5L^2/2+O(L)\)
rs \(3L^2+O(L)\) \(3L^2+O(L)\) \(5L^2/2+O(L)\)
In case only the \(A^{(0)}_n\) are needed (as they have the best convergence properties), the number of divisions in the FS/qd-algorithm can be reduced from \(5L^2/2+O(L)\) to \(2L^2+O(L)\). In any case, we see that the operation count of the rs-algorithm is about 30% more than that of the FS/qd-algorithm. Finally, we observe that when \(u_k=\Delta A_k=A_{k+1}-A_k\) the higher-order \(G\)-transformation reduces to the Shanks transformation, and therefore the rs-and FS/qd-algorithms can be used for computing the \(A^{(j)}_n\) in this case too. Of course, the most famous and efficient implementation of the Shanks transformation is via the \(\varepsilon\)-algorithm of Wynn, which reads \[\label{eq:18} \varepsilon^{(j)}_{k+1}=\varepsilon^{(j+1)}_{k-1}+\frac{1} {\varepsilon^{(j+1)}_{k}-\varepsilon^{(j)}_{k}}, \ \ j,k=0,1,\ldots,\] with the initial values \(\varepsilon^{(j)}_{-1}=0\) and \(\varepsilon^{(j)}_{0}=A_j\) for all \(j\geq 0\). Then \(\varepsilon^{(j)}_{2n}=A^{(j)}_n\) for all \(j\) and \(n\). Thus, given \(A_0,A_1,\ldots,A_{2L}\), we can compute \(\varepsilon^{(j)}_{2n}\) for \(0\leq j+2n\leq 2L\). (In particular, we can compute the diagonal approximations \(\varepsilon^{(j)}_{2n}\), \(n=0,1,\ldots,L,\) that have the best convergence properties.) Since there are \(2L^2+O(L)\) of the \(\varepsilon^{(j)}_k\) to compute, the operation count of this computation is \(4L^2+O(L)\) additions and \(2L^2+O(L)\) divisions and no multiplications. It is seen from the table above that the FS/qd-algorithm compares very favorably with the \(\varepsilon\)-algorithm as well. | {'timestamp': '2017-06-07T02:07:09', 'yymm': '1706', 'arxiv_id': '1706.01786', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01786'} |
null | null |
# Introduction
In the foundations of quantum mechanics, the status of the wavefunction is one of the most debated issues. Associating a wavefunction with each system, QM provides very accurate predictions about the possible measurement outcomes. However, the axioms of QM are silent about the nature of the wavefunction, i.e., these axioms are logically consistent with several different ontological statuses of the wavefunction. The Ontological Models Framework (OMF) introduced by Harrigan and Spekkens, is an attempt to provide a general formalism by which one can evaluate the ontology of the quantum theory and specifically the ontology of the wavefunction in realistic interpretations of QM. Leifer explains that one can choose one of the following three philosophical stands toward the wavefunction's interpretation.
1. Antirealist Epistemic: Wavefunctions are epistemic, and there is not any deeper reality behind them.
2. Realist Epistemic: Wavefunctions are epistemic, but there is an underlying reality.
3. Realist Ontic: There is an underlying reality, and the wavefunctions correspond to the states of reality.
Options one and two are called \(\psi\)-epistemic views. In these views, one interprets the wavefunction as a mathematical object which merely represents the agent's knowledge of the quantum systems. Option three is referred to as the \(\psi\)-ontic view, within which the wavefunction is considered to be a *real* object corresponding to an underlying reality of systems; such an object exists independently of the observer. An important question is whether it is possible in principle to construct a realist \(\psi\)-epistemic interpretation of QM (option two), or whether a realist is compelled to think ontologically about wavefunctions. PBR try to answer this question, and their answer is negative. In their famous article, PBR introduce an operational scenario whose statistics cannot be reproduced by \(\psi\)-epistemic ontological models. PBR concluded that there is no way of having a realist epistemic interpretation of \(\psi\) (option two); thus, the only remaining option for a realist is to choose option three. The logical structure of the PBR's argumentation can be expressed in the following way. There is a contradiction between the statistics of quantum theory (\(P_1\)) and the statistics of \(\psi\)-epistemic ontological models (\(P_2\)) if one accepts a seemingly reasonable assumption about the ontological structure of the ontic states of a composite system, called the "Preparation Independence Postulate\" (\(P_3\)). In general, these statements are going to be discussed within the ontological models framework (\(P_4\)). \[\underbrace{ QSTAT }_{\text{\(P_1\)}}\hspace{2mm}\land\hspace{2mm} \underbrace{ \psi-epistemic }_{\text{\(P_2\)}}\hspace{2mm}\land \underbrace { PIP }_{\text{\(P_3\)}}\hspace{2mm}\land\hspace{2mm} \underbrace{ OMF }_{\text{\(P_4\)}}\hspace{2mm} \Rightarrow\hspace{2mm}\bot\] In confronting the above contradiction, so far, several arguments have been suggested aiming to put doubt on one (or even more than one) of the above propositions. Most of these suggestions can be classified into three categories.
1. Rejection of \(P_1\). In this category, one can argue that maybe QM's predictions are not always true. The idea of having null-measurements (measurements with no outcome), or detection-inefficiency loophole can be considered in this class.
2. Rejection of \(P_2\). Another option is to accept the above contradiction, as a clue of a subtle constraint on accepting the \(\psi\)-epistemic models. PBR themselves claimed that for a realist person, the only remaining option is to be a \(\psi\)-ontologist.
3. Rejection of \(P_3\). Most researchers have criticized the PIP assumption. Arguing that the two subsystems with a common past may share some ontic states in the present or the idea that there may be some holistic properties of a composite system which cannot be reduced to its sub-systems, and several other arguments which can be found in, are examples of such criticisms.
4. Rejection of \(P_4\). Except for the arguments from antirealists, who entirely reject the existence of an underlying reality behind the quantum states, there is no serious criticism of the PBR's theorem which can be classified in this category. However, the argument provided in this study falls into this category.
This research highlights an incorrect conclusion in PBR's arguments, which was made due to inadequacies in the definitions of OMF. To be precise, OMF models the ontology of the preparation procedure, but it does not model the ontology of the measurement device. Such an asymmetric treatment becomes problematic, in scenarios in which measurement devices have a quantum nature. Modifying the OMF definition so that we can insert the ontology of the measurement procedure into our analysis, we can see that PBR's contradiction is not because of the \(\psi\)-epistemic models. Rather, the problem is in the heart of our framework (OMF). I refer this modified version of the OMF as to "MOMF\" (Modified Ontological Models Framework). The order of the paper's sections is as follows. A review of the OMF's definitions and how they relate to the axioms of QM, as well as a precise description of PBR's scenario in OMF, are provided in section two. Pointing out the problem of the \(\psi\)-epistemic ontological models caused by OMF, the mathematical structure of the MOMF formalism followed by a statement of the PBR's theorem within MOMF are discussed in section three. There, We will see how without using any further assumptions, the problem of the \(\psi\)-epistemic models can be solved in MOMF. In conclusion, by reviewing all the mentioned ideas, it will be argued that PBR's finding does not provide any implication for the nature of the wavefunction, rather what they have found shows us that we should be more careful about the way we model the ontology of QM.
# OMF and the PBR Theorem
## Operational QM
Consider an operational \(PM\) scenario, i.e., a scenario in which a system is prepared via a procedure/device \(P\), and then measured via a procedure/device \(M\). In this situation, the interaction between the prepared system and the measurement device results in an outcome for which QM provides us the associated probability of finding that outcome. Figure ([\[PM_EXP\]](#PM_EXP){reference-type="ref" reference="PM_EXP"}) depicts such a scenario. According to QM, we can express the following axioms for describing such a situation.
The idea of OMF is straightforward; As an operational theory, quantum theory associates an operational state \(\rho\) with each preparation procedure. But, since we are not sure whether \(\rho\) is the system's ontic state (\(\lambda\)) or not, we should simulate the ontic level of \(\rho\) by the expression \(pr(\lambda\mid P)\). After this simulation, the probabilities of finding different measurement outcomes will depend on the ontic states of the preparation procedure, expressed by \(pr(m\mid\lambda, M)\). This implies that the outcomes' probabilities depend on the measurement procedure \(M\) and the ontic state \(\lambda\). Finally, these probabilities should be summed over all the relevant ontic states.
The upper graph of figure ([\[on-ep\]](#on-ep){reference-type="ref" reference="on-ep"}) depicts the ontic state space of a \(\psi\)-ontic ontological model, in which, there is no shared \(\lambda\) between two preparation procedures. The lower graph shows a \(\psi\)-epistemic ontological model. In \(\psi\)-ontic models, two operationally distinct states (say \(\ket{x+}\) and \(\ket{z+}\)) are ontologically distinct too. In \(\psi\)-epistemic models, there exists (at least) one ontic state which is compatible with two different operational states. So far, all of our statements about OMF have considered the single-party \(PM\) scenarios (a *single* system is assumed to be prepared and then measured). In this sense, OMF aims to provide the most general ontological structure behind the operational \(PM\) scenario. However, most of the problematic scenarios in foundations of QM are multipartite ones. For instance, the Bell and the PBR scenarios are both about two systems which are going to be joined to create a single composite system.
## The PBR Theorem in OMF
Imagine Alice and Bob are living in two space-like separated regions, and they prepare their quantum states via the preparation procedures \(P_A\) and \(P_B\) such that there is not any operational correlation between these two processes. In the original example of PBR, Alice can choose to prepare her state in either \(\ket{0_A}\) or \(\ket{+_A}\). Similarly, Bob can prepare his state in either \(\ket{0_B}\) or \(\ket{+_B}\) (PBR express these statements, by demanding two copies of the same prepared system). According to figure ([\[on-ep\]](#on-ep){reference-type="ref" reference="on-ep"}), an arbitrary \(\psi\)-epistemic ontological model implies that Alice and Bob have the overlaps \(\Delta_A\) and \(\Delta_B\) in the probability distributions over their ontic state spaces \(\Lambda_{A}\) and \(\Lambda_{B}\). Since these people are living in two space-like separated regions, there is not any operational correlation between their preparation procedures. Hence, according to the operational QM the following relation holds, \[\label{p-op} \psi_{AB} = \psi_A\otimes\psi_B.\] So, there are four possibilities for preparing the composite system AB, \[\psi_{AB}\in \{\psi^1_{AB},\psi^2_{AB},\psi^3_{AB},\psi^4_{AB}\}\] where \(\psi^1_{AB}=\ket{0_A}\otimes\ket{0_B}\), \(\psi^2_{AB}=\ket{0_A} \otimes\ket{+_B}\), \(\psi^3_{AB}=\ket{+_A}\otimes\ket{0_B}\) and \(\psi^4_{AB}=\ket{+_A}\otimes\ket{+_B}\). As it can be seen in figure ([\[com-sup\]](#com-sup){reference-type="ref" reference="com-sup"}), in general, the support of the composite ontic state \(\lambda_{AB}\) is composed of nine mutually exclusive sub-regions. \[\label{similarity} supp(\lambda_{AB}\mid\psi_{AB}) = R_{11}\sqcupdot R_{12}\sqcupdot... \sqcupdot R_{33}\] Also, note that corresponding to each operational state \(\psi^i_{AB}\) there exists a particular region of \(\Lambda_{AB}\), to which the ontic state \(\lambda_{AB}\) will be restricted. For instance, if the composite system has been prepared in \(\psi^1_{AB}=\ket{0_A}\otimes\ket{0_B}\), one can infer that Alice and Bob have prepared their systems in \(\ket{0_A}\) and \(\ket{0_B}\), respectively. In this situation, \(\lambda_{AB}\) will be restricted to the region \(R_{31}\sqcupdot R_{32}\sqcupdot R_{21}\sqcupdot R_{22}\). In general, we have the following four formulas. \[\label{OMF-Supports} \begin{cases} supp(\lambda_{AB}\mid \psi^1_{AB}) = R_{21}\sqcupdot R_{22}\sqcupdot R_{31}\sqcupdot R_{32}\\ supp(\lambda_{AB}\mid \psi^2_{AB}) = R_{11}\sqcupdot R_{12}\sqcupdot R_{21}\sqcupdot R_{22}\\ supp(\lambda_{AB}\mid \psi^3_{AB}) = R_{22}\sqcupdot R_{23}\sqcupdot R_{32}\sqcupdot R_{33}\\ supp(\lambda_{AB}\mid \psi^4_{AB}) = R_{12}\sqcupdot R_{13}\sqcupdot R_{22}\sqcupdot R_{23}\\ \end{cases}\] From the above equations, it is evident that \(R_{22}\) is the only region of \(\Lambda_{AB}\) which is repeating in all of the four equations. This means that \(R_{22}\) is compatible with all of the four possible preparation procedures of \(\psi_{AB}\). While the first assumption of PIP ([\[PIP-In-OMF-1\]](#PIP-In-OMF-1){reference-type="ref" reference="PIP-In-OMF-1"}) implies that \(R_{22}=\Delta_A\times\Delta_B\), its second assumption ([\[PIP-In-OMF-2\]](#PIP-In-OMF-2){reference-type="ref" reference="PIP-In-OMF-2"}) implies that the ontic state \(\lambda_{AB}=(\lambda_A,\lambda_B)\) of each run of the experiment will be chosen from \(R_{22}\) with the non-zero probability \(q=\frac {|\Delta_A\times\Delta_B|}{|supp(\lambda_{AB}\mid\psi_{AB})|}>0\). As we will see in the next pages, \(\psi\)-epistemic models cannot reproduce PBR statistics because of the appearance of \(R_{22}\). In other words, \(R_{22}\) is a problematic region in \(\psi\)-epistemic models. So far, we have reviewed many details of PBR preparation procedure. Before explicitly expressing the main theorem, PBR measurement procedure should also be known. PBR define their measurement procedure to be \(M^{AB}=\{ M^{AB}_1,M^{AB}_2,M^{AB}_3,M^{AB}_4 \}\), which is defined as follows. \[\begin{aligned} M^{AB}_1=\frac{1}{\sqrt{2}}(\Ket{0_A}\Ket{1_B}+\Ket{1_A}\Ket{0_B})\\ M^{AB}_2=\frac{1}{\sqrt{2}}(\Ket{0_A}\Ket{-_B}+\Ket{1_A}\Ket{+_B})\\ M^{AB}_3=\frac{1}{\sqrt{2}}(\Ket{+_A}\Ket{1_B}+\Ket{-_A}\Ket{0_B})\\ M^{AB}_4=\frac{1}{\sqrt{2}}(\Ket{+_A}\Ket{-_B}+\Ket{-_A}\Ket{+_B}) \end{aligned}\] It can be verified that elements of the above set constitute a complete, two-by-two orthogonal basis corresponding to four possible outcomes of a measurement device \(M^{AB}\). This measurement procedure is entangled, meaning that the elements of this set cannot be separated into two parts, \(M^{AB}_i \neq (M^A_i, M^B_i)\).
# MOMF and the PBR Theorem
A careful investigation of the given proof of the PBR theorem within OMF suggests that the problem of PBR's theorem is neither due to the nature of \(\psi\) nor due to the \(\psi\)-epistemic models. Rather, it seems that OMF is not capable enough in describing the measurement procedures performed on the *entangled* basis. According to the OMF's definition, the response functions depend on the *ontic* states of the preparation device \(\{\lambda\}\) and the *operational* states of the measurement apparatus \(\{M_m\}\). This means that, although OMF considers the ontology of the preparation procedure, asymmetrically it does not consider the ontology of the measurement device. Such a treatment becomes problematic, especially in scenarios in which measurement devices have a quantum nature. In such circumstances, the preparation, and the measurement devices create a composite system whose behavior depends on the ontic states of both \(P\) and \(M\). Let us denote the corresponding ontic states of \(P\) and \(M\) by \(\lambda^P\) and \(\lambda^M\), respectively. Needless to say, the measurement procedure of PBR's scenario is an entangled one implying that the quantum nature of \(M\) is crucially important. Allowing the response functions to depend on \(\lambda^P\) and \(\lambda^M\), in the following subsections, we will find that there is no contradiction between the PBR operational statistics and the predictions of the \(\psi\)-epistemic models.
## Modified Ontological Models Framework
The above definition is too general and even useless in most of the cases. However, putting some reasonable constraints on this definition converts it into a more useful form. For instance, one can reasonably assume that \(P\) and \(M\) are operationally uncorrelated. Then, by exploiting an auxiliary assumption which mirrors the operational independence of \(P\) and \(M\) to the ontological level, the following statement can be obtained.
The ideas of the PM-Separability assumption and the PIP assumption are very similar; in the absence of operational correlations between two events, one may assume that these events are ontologically uncorrelated too. As will be explained later, such auxiliary assumptions are not necessarily reasonable. However, the purpose of this paper is to show that even if we accept PBR's assumption, their conclusion does not follow.
Equation ([\[two-density\]](#two-density){reference-type="ref" reference="two-density"}) shows the reasonableness of the previous statements; that for simulating the ontology of an operational scenario, one needs both ontic states \(\lambda^P\) and \(\lambda^M\). Although MOMF approaches symmetrically toward the preparation and measurement procedure, there is a conceptual difference between the forms of \(\Lambda^P\) and \(\Lambda^M\). In principle, it is possible for different preparation methods to share some ontic states in their supports (this is the definition of \(\psi\)-epistemic models). However, in the measurement ontic state space \(\Lambda^M\), this statement cannot hold, meaning that different regions of the ontic space (corresponding to different outcomes) are disjoint. Note that the outcomes of a measurement device correspond to different orthogonal POVM. Hence, from the operational points of view, the probability of being in an (operational) state which is compatible with two orthogonal POVMs is equal to zero. This operational claim can be confirmed by our daily experience, where we have never seen a measurement's pointer being in two different outcomes at the same time. From the ontological points of view, one should note that \(\Lambda^M\) must be composed of several non-overlapping regions, corresponding to several measurement outcomes. Otherwise, we could have a set of shared ontic states \(\lambda^M\), each of which being (ontologically) compatible with (operationally) incompatible states. Hence, via mirroring the observations coming from the operational level of QM into the ontological level, we have the following assumption. Figure ([\[measurement\]](#measurement){reference-type="ref" reference="measurement"}) depicts this situation.
## The PBR Theorem in MOMF
Translating the details of the PBR's operational scenario in MOMF, now, we can see the reason of why the PBR's conclusion is not necessarily correct. First of all, let us explicitly express all the constraints on MOMF entailed by PBR's scenario.
1. We assume that the \(PM\) separability assumption is valid, \[\Lambda^{PM}=\Lambda^P\times\Lambda^M\equiv\lambda^{PM}=(\lambda^P,\lambda^M)\] \[pr(\lambda^{PM}\mid PM)=pr(\lambda^P\mid P)pr(\lambda^M\mid M)\]
2. According to PBR's scenario, the preparation procedures are operationally separable, \[P_{AB}=(P_A, P_B)\] \[\rho_{AB}=\rho_A\otimes \rho_B\]
3. According to PBR's scenario, the measurement procedures are entangled, \[M^{AB}_i\neq(M^A_i, M^B_i)\].
4. We assume that the PIP assumption is valid, \[\Lambda^{P}_{AB}=\Lambda^P_A\times\Lambda^P_B\equiv\lambda^{P}_{AB}=(\lambda^P_A,\lambda^P_B)\] \[pr(\lambda^P_{AB}\mid P_{AB})=pr(\lambda^P_A\mid P_A)pr(\lambda^P_B\mid P_B)\]
5. Our model is \(\psi\)-epistemic; Some (preparation) ontic states come from a common region which is compatible with all of the four possible preparation procedures, \[(\lambda^P_{A},\lambda^P_{B})\in\Delta^P_A\times\Delta^P_B\]
6. The support of the measurement ontic space \(\Lambda^M_{AB}\) consists of four disjoint regions corresponding to four measurement outcomes, \[\label{pbr-measurement}\lambda^M_{AB}\in\Delta^M_{AB}=\Delta^M_1\sqcupdot\Delta^M_2\sqcupdot\Delta^M_3\sqcupdot\Delta^M_4\]
**Step 1.** In this step, by combing the above constraints on MOMF definition, we derive further required equations. Making the PIP and the PM separability assumptions, the total ontic state space \(\Lambda^{PM}_{AB}\) can be obtained as follows, \[\label{total-space} \Lambda^{PM}_{AB}=\Lambda^P_{AB}\times\Lambda^M_{AB}=\Lambda^P_{A}\times\Lambda^P_{B}\times\Lambda^M_{AB}\] Exactly similar to ([\[similarity\]](#similarity){reference-type="ref" reference="similarity"}), the support of the preparation ontic state, \(supp(\lambda^P_{AB})\), is composed nine mutually exclusive sub-regions. \[supp(\lambda^{P}_{AB}\mid P_{AB}) = R^P_{11}\sqcupdot R^P_{12}\sqcupdot... \sqcupdot R^P_{33}\] On the other hand, the support of the measurement ontic state, \(supp(\lambda^{M}_{AB}\), is composed of four mutually exclusive sub-regions. Placing ([\[pbr-measurement\]](#pbr-measurement){reference-type="ref" reference="pbr-measurement"}) in ([\[total-space\]](#total-space){reference-type="ref" reference="total-space"}), it can be understood that the support of the total ontic state, \(supp(\lambda^{PM}_{AB}\), is composed of thirty-six mutually exclusive sub-regions. \[\label{supports} supp(\lambda^{PM}_{AB}\mid P_{AB}, M^{AB}) = supp(\lambda^{P}_{AB}\mid P_{AB})\times supp(\lambda^{M}_{AB}\mid M^{AB})=\] \[(R^P_{11}\sqcupdot...\sqcupdot R^P_{33})\times (\Delta^M_1\sqcupdot...\sqcupdot\Delta^M_4)\] We have already seen that the problem of the \(\psi\)-epistemic models is the existence of the overlap region \(R^P_{22}\), and not the other regions. Hence, to assess the ability of a \(\psi\)-epistemic MOMF model in reproducing PBR statistics, it suffices to check this ability for the states belonging to this problematic region, \(\lambda^{PM}_{AB} \in R^P_{22}\times\Delta^M_{AB}\). In the total ontic state space \(\Lambda^{PM}_{AB}\), there are four sub-regions being compatible with \(R^P_{22} = \Delta^P_A\times\Delta^P_A\). Let us denote them by \(R_1\), \(R_2\), \(R_3\) and \(R_4\), where \(R_i = R^P_{22}\times\Delta^M_{i}\). Figure ([\[three-dim\]](#three-dim){reference-type="ref" reference="three-dim"}) provides a three dimensional representation of these regions. Note that these regions are disjoint, and their union constructs the problematic region \(R^P_{22}\times\Delta^M_{AB}\), \[\Delta^P_{A}\times\Delta^P_{B}\times\Delta^M_{AB}= R_1\sqcupdot R_2\sqcupdot R_3\sqcupdot R_4\] Finally, we have the following equation for computing the probability of finding different outcomes. \[\label{combine} pr(m^{AB} \mid P_{AB}, M^{AB}) =\] \[\resizebox{1\hsize}{!}{\(\int_{\Lambda^P_A}\int_{\Lambda^P_B}\int_{\Lambda^M_{AB}}d\lambda^P_A d\lambda^P_B d\lambda^M_{AB}pr(\lambda^P_A \mid P_A) pr(\lambda^P_B \mid P_B)pr(\lambda^M_{AB} \mid M^{AB}) pr(m^{AB}\mid\lambda^P_A, \lambda^P_B, \lambda^M_{AB})\)}\] **Step 2.** In this step, PBR's observations will be expressed within MOMF. **Observation 1**: Our measurement device can select the outcomes which are non orthogonal to the prepared state. So the probability of choosing an outcome which is orthogonal to the prepared state is zero. This observation leads to the four following equations on the operational level, \[\label{op-set}\textup{operational level} \begin{cases} pr(m^{AB}=1\mid \psi_{AB}=\ket{0_A 0_B},M^{AB})=Tr(M^{AB}_1 \rho^{1}_{AB})=0\\ pr(m^{AB}=2\mid \psi_{AB}=\ket{0_A +_B},M^{AB})=Tr(M^{AB}_2 \rho^{2}_{AB})=0\\ pr(m^{AB}=3\mid \psi_{AB}=\ket{+_A 0_B},M^{AB})=Tr(M^{AB}_3 \rho^{3}_{AB})=0\\ pr(m^{AB}=4\mid \psi_{AB}=\ket{+_A +_B},M^{AB})=Tr(M^{AB}_4 \rho^{4}_{AB})=0\\ \end{cases}\] Combining the above results with equations ([\[supports\]](#supports){reference-type="ref" reference="supports"}) and ([\[combine\]](#combine){reference-type="ref" reference="combine"}), one can convert this set of four operational equations into the following set of equations at the ontic (MOMF) level of QM. \[\label{obser1}\textup{ontic level} \begin{cases} pr(m^{AB}=1\mid\lambda^P_A ,\lambda^P_B,\lambda^M_{AB})=0\hspace{2mm} when \hspace{2mm}\lambda^{PM}_{AB}\in (R^P_{21}\sqcupdot \textcolor{red}{R^P_{22}}\sqcupdot R^P_{31}\sqcupdot R^P_{32})\times (\Delta^M_1)\\ pr(m^{AB}=2\mid\lambda^P_A ,\lambda^P_B,\lambda^M_{AB})=0\hspace{2mm} when \hspace{2mm}\lambda^{PM}_{AB}\in (R^P_{11}\sqcupdot R^P_{12}\sqcupdot R^P_{21}\sqcupdot \textcolor{red}{R^P_{22}})\times (\Delta^M_2)\\ pr(m^{AB}=3\mid\lambda^P_A ,\lambda^P_B,\lambda^M_{AB})=0\hspace{2mm} when \hspace{2mm}\lambda^{PM}_{AB}\in (\textcolor{red}{R^P_{22}}\sqcupdot R^P_{23}\sqcupdot R^P_{32}\sqcupdot R^P_{33})\times (\Delta^M_3)\\ pr(m^{AB}=4\mid\lambda^P_A ,\lambda^P_B,\lambda^M_{AB})=0\hspace{2mm} when \hspace{2mm}\lambda^{PM}_{AB}\in (R^P_{12}\sqcupdot R^P_{23}\sqcupdot \textcolor{red}{R^P_{22}}\sqcupdot R^P_{23})\times (\Delta^M_4)\\ \end{cases}\] **Observation 2**: According to equation ([\[sharte mohem\]](#sharte mohem){reference-type="ref" reference="sharte mohem"}), regardless of which quantum state has been prepared, the sum of the response functions must be equal to one. \[\label{obser2} \Sigma^4_{i=1} pr(m^{AB}=i\mid\lambda^P_{A},\lambda^P_{B},\lambda^M_{AB})\] **Step 3.** Fortunately, there is no contradiction between the implications of the first observation, in ([\[obser1\]](#obser1){reference-type="ref" reference="obser1"}), and the consequences of the second observation, in ([\[obser2\]](#obser2){reference-type="ref" reference="obser2"}): The ontic states \(\lambda^{PM}_{AB}=(\lambda^P_A,\lambda^P_B,\lambda^M_{AB})\) in equations of the set ([\[obser1\]](#obser1){reference-type="ref" reference="obser1"}) are not the same. Rather, each triplet belongs to one specific sub-region of the total ontic space which is disjoint from the others. Hence, if the response function of our model assigns zero to one triplet, the other three triplets become non-zero. And exactly these three triplets can reproduce the first observation ([\[obser1\]](#obser1){reference-type="ref" reference="obser1"}). To be precise, different parts of the measurement ontic state space are completely disjoint from each other. So, we have \[|\Delta^M_i \cap\Delta^M_{j \neq i}|= 0\] Hence, supports of the ontic state in four equations ([\[obser1\]](#obser1){reference-type="ref" reference="obser1"}) do not share any common ontic state, \[|(\textcolor{red}{R^P_{22}}\times\Delta^M_i)\cap (\textcolor{red}{R^P_{22}}\times\Delta^M_{j\neq i})|= 0\]
# Conclusion
First, I investigated the PBR theorem and its underlying framework, OMF. Pointing to a problem inside the OMF definition, I defined a modified version of OMF, MOMF. Translating the PBR's scenario in MOMF, I showed that PBR's claim in rejecting the epistemic view is not because of the nature of the wavefunction. Rather, (I argued that) the reason of why PBR have found such a result lies in the OMF definition. That this, the PBR theorem has vanished within MOMF. Inserting the measurement ontic state was the key idea of this paper. This analysis suggests that the importance of PBR's work is not because of their conclusion about the nature of the wavefunction. Rather, PBR's finding reveals a significant deficiency of OMF in not considering the ontic level of measurement devices. It is important to note that, the arguments of this paper do not lead to any conclusion about the nature of wavefunction. However, the paper reveals that *PBR's arguments cannot* reject the epistemic view of the wavefunction. | {'timestamp': '2017-06-07T02:07:54', 'yymm': '1706', 'arxiv_id': '1706.01819', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01819'} |
# Introduction
An emergent class of transient objects defined by unusually strong calcium line emissions that develop in optical spectra months after explosion has garnered considerable attention in the last decade. These have been referred to as "Ca-rich,", SN 2005E-like, and "Ca-rich gap" transients. Here we adopt the broad classification of "Ca-rich transient." During the early photospheric stages, Ca-rich transients have distinct He lines in their spectra and fall within the supernova (SN) Type Ib classification (see @Filippenko97 [@Gal-Yam16], and @Parrent16 for reviews of the SN classification scheme). However, Ca-rich transients are less luminous than normal SNe Ib and reach optically thin (i.e., "nebular") stages within two months post-maximum light. During this stage they exhibit a large emission line ratio \[\]/\[\] \(>\) 2, which suggests that the total calcium synthesized is greater than that of normal Type I supernovae, perhaps by factor of \(5-10\). Estimates of their rates range from 2% to 12% of SNe Ia. The progenitors of Ca-rich transients are not confidently known. Stars with initial masses in the range of \(8-12\;\rm M_{\odot}\) have been suggested, but unlike the majority of core-collapse SNe, many Ca-rich transients have been hosted by early-type galaxies lacking obvious massive star populations. Locations within dwarf galaxies or globular clusters is a possible resolution to this inconsistency, but close examinations of explosion sites have thus far failed to uncover any sign of in situ star formation. recently proposed that the Galactic supernova remnant RCW 86 formed from a core-collapse Ca-rich supernova that polluted the progenitor star's solar-type binary companion that evolved through a common envelope phase. Ca-rich transients are often found at large distances away from their host galaxies (as high as 150 kpc; @Perets10 [@Kasliwal12; @Perets14; @Foley15]). This intriguing property, combined with the lack of nearby stellar populations, has supported the notion that the progenitor systems have traveled significant distances before exploding, perhaps via interaction with a super-massive black hole (BH). The relatively large delay-time distribution required to travel large distances favors an older white dwarf (WD) population that also contributes to SNe Ia. Many explosion channels have been proposed, including helium detonations occurring on a helium-accreting WD, tidal disruptions of low mass WDs by neutron stars or stellar mass BHs, and tidal detonations of WDs caused by random flyby encounters with an intermediate mass BH (IMBH) in dwarf galaxies or globular clusters. Though the class has generally been treated uniformly, suspicions of heterogeneity in the progenitor system have been voiced. Here we present and analyze the multi-wavelength data set of a supernova exhibiting a revealing mix of properties that bridge Ca-rich transients and core-collapse SNe. Our observations are provided in sections [2](#sec:iPTF15eqv){reference-type="ref" reference="sec:iPTF15eqv"} and [3](#sec:observations){reference-type="ref" reference="sec:observations"}, our results in section [4](#sec:results){reference-type="ref" reference="sec:results"}, and a detailed comparative analysis of the entire class of Ca-rich transients in section [5](#sec:discussion){reference-type="ref" reference="sec:discussion"}. In section [6](#sec:conclusions){reference-type="ref" reference="sec:conclusions"} we review the ramifications of our conclusion that the progenitor system of the Ca-rich transient iPTF15eqv was most likely a massive star, and describe future observations and modeling required to understand the wide-ranging diversity encompassed in the Ca-rich transient observational classification.
# iPTF15eqv {#sec:iPTF15eqv}
On 2015 September 27.81 (all dates UT) we reported discovery of PSN J10520833+3256394 in NGC 3430 to the Central Bureau for Astronomical Telegrams "Transient Object Followup Reports" page.[^1] Months later an independent discovery was made by the intermediate Palomar Transient Factory (iPTF) on 2015 December 6 and designated iPTF15eqv, which we adopt here. Spectra obtained on 2015 Dec 07.53 by led them to classify iPTF15eqv as a Type IIb/Ib supernova in the nebular phase. Description of the spectrum noted weak continuum emission superposed by strong \[\] \(\lambda\lambda\)`<!-- -->`{=html}7291, 7324 and weak H\(\alpha\) emission lines. The lack of equally strong \[\] \(\lambda\lambda\)`<!-- -->`{=html}6300, 6364 emission is unusual for a SN IIb/Ib and prompted our group to obtain optical spectra that confirmed the non-standard nature of the emissions.[^2] A multi-wavelength follow up campaign supported by radio observations with the Karl G. Jansky Very Large Array (VLA) and X-ray observations with Chandra X-ray Observatory (CXO) was initiated. Distance estimates to the host galaxy NGC 3430 range from 15.4 Mpc to 32.9 Mpc. We adopt an estimate of 30.4 Mpc with distance modulus \(\mu = 32.41 \pm 0.43\). NGC 3430 is classified as a Type SBc spiral by, and also hosted the Type II SN 2004ez. Extinction toward iPTF15eqv appears to be minimal. Foreground extinction due to the Milky Way is \(E(B-V)_{mw} = 0.021\) mag, there is no obvious suppression of flux at short wavelengths, and no conspicuous D absorption is observed in our optical spectra (see section [3.2](#sec:spectra){reference-type="ref" reference="sec:spectra"} for details). Thus, no correction for extinction has been adopted in this paper to the photometry or spectra. The one exception is an estimate of the metallicity of the host system (section [4.4](#sec:metallicity){reference-type="ref" reference="sec:metallicity"}) that uses the relative strengths of narrow emission lines associated with the environment around iPTF15eqv. In Figure [\[fig:fchart\]](#fig:fchart){reference-type="ref" reference="fig:fchart"}, we show a finding chart of iPTF15eqv made from images of the region obtained with the 6.5-m MMT telescope in combination with the MMTCam instrument[^3] in \(g^{\prime}r^{\prime}i^{\prime}\)-band filters. This composite image highlights the location of iPTF15eqv, which lies along an exterior spiral arm approximately 44\(^{\prime\prime}\) (\(\approx 6.5\) kpc) from the host galaxy nucleus.
# Observations {#sec:observations}
## Optical photometry
Unfiltered images of iPTF15eqv and its surrounding environment were obtained at Itagaki Observatory using the 0.60-m \(f/5.7\), 0.50-m \(f/6.0\), and 0.35-m \(f/11.0\) telescopes mounted with a Bitran-CCD (Kodak KAF 1001E) detector[^4] between 2015 September 27 and 2015 December 18. We also obtained three epochs of MMTCam \(r'i'\)-band photometry and one epoch of \(g'\)-band photometry between 2016 February 06 and 2016 June 08. A log of the observations is provided in Table 1, and a plot of the final photometry in Figure [\[fig:lc\]](#fig:lc){reference-type="ref" reference="fig:lc"}. Unless otherwise noted, uncertainties are quoted at the 1\(\sigma\) confidence level. Processing of the images was completed using standard routines in the Image Reduction and Analysis Facility (IRAF).[^5] This included bias, flat-field, and dark frame corrections, when available. Multiple images in a single night were stacked. We performed point-spread function (PSF) photometry on iPTF15eqv and nearby field stars. Absolute calibration was carried out using Sloan Digital Sky Survey (SDSS) observations of field stars. Approximate unfiltered magnitudes for the field stars were calculated by performing synthetic photometry on the SDSS \(u'g'r'i'z'\) spectral energy distribution using the typical spectral response achieved by the KAF 1001E detector in use at Itagaki Observatory. In addition, we subtract the flux associated with an \(\sim\)`<!-- -->`{=html}21 mag stellar knot located 1.5\(^{\prime\prime}\) from iPTF15eqv from the unfiltered magnitudes measured from the Itagaki Observatory images. This stellar knot is well resolved and separated from the SN in our MMTCam observations, but would be included within the aperture used on the Itagaki Observatory images (typical resolution \(\approx\)`<!-- -->`{=html}4\(^{\prime\prime}\)). This correction is \(\lesssim\)`<!-- -->`{=html}0.1 mag for all epochs. In order to directly compare the early unfiltered light curve to our late-time MMTCam observations, we performed synthetic photometry on a spectrum of iPTF15eqv obtained on 2016 January 18 (see section [3.2](#sec:spectra){reference-type="ref" reference="sec:spectra"}) to investigate the relationship between the unfiltered magnitudes of the KAF 1001E detector and standard bandpasses. We verified the relationship using spectra of the Ca-rich transient SN 2005E. For transients with large \(r'\)-\(i'\) colors (as is observed for iPTF15eqv at late times due to the large \[\] feature) we find an approximate relationship between unfiltered and \(r'\)-band magnitudes which ranges from \(r'-\)unf = 0.0 mag for \(r'-i'\) \(= 0.85\) mag to \(r'-\)unf \(= 0.3\) mag for \(r'-i'\) \(= 1.45\) mag. Applying these shifts to our MMTCam data we obtain approximate unfiltered magnitudes for iPTF15eqv at late epochs that can be compared to the earlier phase data from Itagaki Observatory. These data are shown as open squares in Figure [\[fig:lc\]](#fig:lc){reference-type="ref" reference="fig:lc"}. We also plot unfiltered, \(r'\)-and \(i'\)-band magnitudes obtained from our synthetic photometry. We do not attempt to transform the unfiltered magnitudes obtained from the Itagaki Observatory images to a standard bandpass since no spectral or color information is available at these epochs.
## Optical spectroscopy {#sec:spectra}
Low-resolution optical spectra of iPTF15eqv were obtained using the MMT telescope in combination with the Blue Channel instrument, and the Hiltner 2.4 m telescope in combination with the Ohio State Multi-Object Spectrograph (OSMOS; @osmospaper) instrument and R4K detector at MDM observatory. Conditions were photometric for most observations. Standard stars from and were observed for spectrophotometric calibration and to remove telluric features. A log of all optical spectroscopy is provided in Table 2. The Blue Channel observations used the B300 grating in combination with a 1\(^{\prime\prime}\) slit width yielding spectra spanning a wavelength window of \(3400-8500\) Å and with resolution \(R \sim 740\). OSMOS observations used a 1.2 slit width with the VPH-red grism and an OG530 order-blocking filter, resulting in spectra spanning a wavelength window of approximately \(5500-10500\) Å with resolution \(R \sim 1200\). Standard procedures to bias-correct, flat-field, and flux calibrate the data were followed using the IRAF/PYRAF software[^6] and our own IDL routines. Narrow and unresolved emission lines of \[\] 3727, \[\] 4959, 5007, H\(\alpha\), H\(\beta\), and \[\] 6716, 6731 attributable to a coincident region were used to determine a recession velocity of 1494 km s\(^{-1}\), which was removed from all spectra. A plot of the observations labeling prominent emission features is presented in Figure [\[fig:spectra\]](#fig:spectra){reference-type="ref" reference="fig:spectra"}.
## Near-infrared spectroscopy and photometry
A single epoch of low-resolution, near-infrared (NIR) spectra spanning \(1-2.4\,\mu\)m was obtained on 2016 March 23 with MMT using the MMT and Magellan Infrared Spectrograph (MMIRS) instrument. Observations were made using a 0.6\(^{\prime\prime}\) slit width in three configurations: zJ filter (spanning 0.95-1.50 \(\mu\)m) + J grism (\(R \sim 2000\)), H filter (spanning 1.50-1.79 \(\mu\)m) + H3000 grism (\(R \sim 3000\)), and Kspec filter (spanning 1.95-2.45\(\mu\)m) + K3000 grism (\(R \sim 3000\)). Exposure times were \(4\times 300\) sec for each configuration, and the slit was dithered between exposures. Data were processed using the standard pipeline, to develop wavelength calibrated 2D frames from which 1D extractions were made. A plot of these data combining all three individual spectra is shown in Figure [\[fig:nirspec\]](#fig:nirspec){reference-type="ref" reference="fig:nirspec"}. Line identifications are discussed in section [4.2](#sec:identify){reference-type="ref" reference="sec:identify"}. Acquisition images obtained in J-band were used to flux calibrate our NIR spectra. We performed aperture photometry on a stacked image made from pipeline-processed individual images, choosing an aperture to encompass all of the SN flux. 2MASS magnitudes of field stars provided a means of absolute calibration. We estimate the J-band magnitude of iPTF15eqv on 2016 March 23 to be \(19.77 \pm 0.16\) mag.
## Chandra X-ray observations {#sec:x-ray}
We obtained deep X-ray observations of iPTF15eqv with CXO on 2016 January 4 under our approved program to explore stellar deaths at high energies and radio wavelengths (observation ID 16821, PI: Margutti). CXO data have been reduced with the CIAO software v4.8 and corresponding calibration files. No X-ray source is detected at the location of iPTF15eqv with a total exposure time of 9.9 ks. The inferred \(3\,\sigma\) count-rate limit is \(4\times 10^{-4}\,\rm{c\,s^{-1}}\) (0.5-8 kev). The neutral hydrogen column density in the direction of iPTF15eqv is \(1.84\times 10^{20}\,\rm{cm^{-2}}\). For an assumed non-thermal spectrum with index \(\Gamma=2\), we derive an unabsorbed (absorbed) flux limit of \(4.8\times 10^{-15}\,\rm{erg\,s^{-1}cm^{-2}}\) (\(4.5\times 10^{-15}\,\rm{erg\,s^{-1}cm^{-2}}\)) in the 0.3-10 keV energy range, corresponding to \(5.3\times 10^{38}\,\rm{erg\,s^{-1}}\) at the distance of iPTF15eqv.
## Jansky Very Large Array radio observations
We obtained five epochs of radio observations of iPTF15eqv at frequencies ranging from 4.9 to 16.0 GHz with the VLA (PI: Margutti, program SG0534). All observations were taken in the standard continuum observing mode with a bandwidth of \(\rm 16 ~IF \times 64 ~channel/IF \times 2 ~MHz/channel\). The VLA changed configurations at various stages during these observations. During the reduction we split the data in two side bands (\(\rm 8 ~IF\) each) of approximately 1 GHz each. We used radio source 3C286 for flux calibration, and calibrator J1018+3542 for phase referencing. Data were reduced using standard packages within the Astronomical Image Processing System (AIPS). No radio emission was detected from iPTF15eqv in any of these observations resulting in deep flux density limits. The results have been summarized in Table 3.
# Results {#sec:results}
## Light curve evolution {#sec:lc}
We have utilized all photometric data of iPTF15eqv to compare it to the light curves of selected Ca-rich transients. Although our unfiltered photometry does not strictly match \(R\)-band photometry, it does provide a reasonable means of comparing the slope and absolute luminosity of the light curve. As shown in Figure [\[fig:ca-rich-lc\]](#fig:ca-rich-lc){reference-type="ref" reference="fig:ca-rich-lc"}, iPTF15eqv was discovered at its peak observed unfiltered absolute magnitude of \(-15.4\), which is in the range of Ca-rich objects (\(R \sim-15\) to \(-16.5\) mag; @Kasliwal12 [@Valenti14]). The decline of the light curve of iPTF15eqv at a rate of \(\approx 0.017\) mag day\(^{-1}\) is shallower than those of other Ca-rich transients at similar epochs. If we assume that iPTF15eqv was discovered around maximum light and shares the typical rise time of Ca-rich transients (12\(-\)`<!-- -->`{=html}15 days; @Kasliwal12), then the explosion date was approximately 2015 September 12. We refer to this as \(t_{\rm exp}(s1)\), and consider it a plausible scenario since the first reported spectrum of iPTF15eqv showing strong \[\] emissions by was obtained 70 days after discovery, which is within the time frame that Ca-rich transients exhibit pronounced Ca emissions. However, because we do not have firm observational constraints on the explosion date, the light curve of iPTF15eqv may have peaked many weeks or months before the discovery date. To investigate this alternative scenario, we attempted to match the light curve of iPTF15eqv to those of SN Ib/c. By shifting the light curve of iPTF15eqv \(\approx 50\) days relative to the date of discovery, the overall slope and luminosity is comparable to the late-time linear decay rates found in other SN Ib/c and consistent with expectations from \(^{56}\)Co-decay (Figure [\[fig:ca-rich-lc\]](#fig:ca-rich-lc){reference-type="ref" reference="fig:ca-rich-lc"}). A particularly good match was found with the light curve of SN 2011dh (@Ergon15). If this analogous connection to SN 2011dh is valid, then the first spectrum of iPTF15eqv was obtained some 120 days after the light curve peak that may have have been as luminous as \(-17.5\) mag. We refer to the explosion date 2015 July 24 of this alternative scenario as \(t_{\rm exp}(s2)\). iPTF15eqv most likely exploded in the time interval \(t_{\rm exp}(s1) > t > t_{\rm exp}(s2)\), and peaked in \(R\)-band absolute magnitude between approximately \(-17.5\) and \(-15\). If iPTF15eqv was, indeed, discovered after maximum light, then its light curve evolution can be approximately described by the nebular light curve model of, which is based on the instantaneous rate of energy deposition from the \(^{56}\)Ni \(\rightarrow\) \(^{56}\)Co \(\rightarrow\) \(^{56}\)Fe decay chain. We investigated the range of viable explosion parameters for iPTF15eqv in this scenario. Model parameters include the total nickel mass, M\(_{\rm{Ni}}\), and a constant, \(F\), which describes the incomplete trapping of gamma-rays. \(F\) is a function of the total ejecta mass, M\(_{\rm ej}\), explosion kinetic energy, E\(_{\rm{K}}\), and ejecta density distribution, and we adopt the parameterization of such that \(F \propto\) M\(_{\rm{ej}}\)/\(\sqrt{E_{\rm{K}}}\) (also see @Clocchiatti97). The luminosity of iPTF15eqv is estimated from the unfiltered light curve, assuming zero bolometric correction. The range in mass of \(^{56}\)Ni required to reproduce the luminosity of iPTF15eqv is degenerate with an adopted explosion date. The slope of the light curve rules out reasonable fits bracketing \(t_{\rm exp}(s1)\). However, for explosions bracketing \(t_{\rm exp}(s2)\) (i.e., 40 \(-\) 60 days prior to discovery), we obtain reasonable fits and constrain M\(_{\rm{Ni}}\) to be between \(\approx 0.04-0.07\) M\(_{\odot}\) (Figure [\[fig:explosionparams\]](#fig:explosionparams){reference-type="ref" reference="fig:explosionparams"}). In order to match the observed decay rate with this model, we require that (M\(_{\rm{ej}}\)/M\(_\odot\))(E\(_{\rm{K}}\)/10\(^{51}\) erg)\(^{-0.5}\) \(=\) 3.0 \(\pm\) 0.1. Assuming an average bulk explosion velocity at peak for Ca-rich transients of \((7-9) \times 10^3\) km s\(^{-1}\) we estimate M\(_{\rm{ej}} \approx 2-4\) M\(_{\odot}\) and E\(_{\rm{K}} \approx (0.8-2) \times 10^{51}\) erg. These explosion parameters are comparable to those found for stripped-envelope SNe, including SN 2011dh.
## Emission line identification, Doppler velocities, and evolution {#sec:identify}
The dominant emission feature of iPTF15eqv over all observed epochs is \[\] \(\lambda\lambda\) 7291, 7324 (Figure [\[fig:spectra\]](#fig:spectra){reference-type="ref" reference="fig:spectra"}). The NIR triplet \(\lambda\lambda\) 8498, 8542, and 8662 is strong on day 95, but has faded by day 158. The critical density for the \[\] doublet is \(\sim 3 \times 10^6~\rm cm^{-3}\) from CHIANTI v7, so this threshold may have been reached during the observed epochs, but the ratio also depends on temperature. Also unambiguously detected is \[\] \(\lambda\lambda\)`<!-- -->`{=html}6300, 6364. The ratio of \[\]/\[\] \(\approx 10\) exhibits little change over the entire observing period. We attribute emission centered around 5890 Å to a blend of and 5876. The central wavelength of the observed distribution favors as the dominant ion, but helium is inferred to be a potential contributor because Ca-rich transients show helium in early spectra and there is evidence of the lines in the NIR. We identify a feature centered near 6560 Å as H\(\alpha\) or \] \(\lambda\)`<!-- -->`{=html}6572. The central wavelength of the feature favors H\(\alpha\). However, the lack of H\(\beta\) emission, the extremely strong \[\] lines, the large \] 6572 collision strength and the presence of \[\] and \[\] in the NIR spectrum all suggest that \] 6572 may be strong. \[\] \(\lambda\lambda\)`<!-- -->`{=html}6548, 6583 may also contribute to the profile. In the day 199 spectrum a minor emission peak is observed around 4550 Å. We attribute this to blueshifted \] 4571. An alternative identification of \(\lambda\)`<!-- -->`{=html}4589 seems unlikely since its stronger companion line \(\lambda\)`<!-- -->`{=html}7725 is not detected. Also seen in the day 199 spectrum is a "plateau" of emission between \(4000-5600\) Å. This emission is not uncommon in SNe Ib/c (see, e.g., @Milisavljevic13-12au), and due largely to iron-peak elements. We show the emission line profile distributions for a variety of ions in Figure [\[fig:lineprofiles\]](#fig:lineprofiles){reference-type="ref" reference="fig:lineprofiles"}. The \[\] lines exhibit the largest velocities. With respect to the \(\lambda\)`<!-- -->`{=html}7291 line measuring toward short wavelengths, and the \(\lambda\)`<!-- -->`{=html}7324 line measuring toward long wavelengths, the emission is above the continuum level for velocities up to and potentially exceeding \(\sim 10^{4}\) km s\(^{-1}\). The profile has a central absorption notch centered around \(+750\) km s\(^{-1}\) situated between two peaks at velocities of \(-1700\) km s\(^{-1}\) and \(+2600\) km s\(^{-1}\). The blue/red ratio between peaks decreases from 3:1 to 3:2 over the epochs observed. The central dip in \[\] 7291, 7324 does not correspond to the separation of the lines (1400 \(\rm km~s^{-1}\)) and must be due to geometry or optical depth. On the other hand, the Ca triplet on day 95 shows a similar dip, but it is redshifted by \(\approx 4000\) km s\(^{-1}\) and the peaks correspond to the separation between the two strongest lines of the triplet. The triplet is almost certainly optically thick, and radiative transfer must play an important role. The \[\], / blend, and H\(\alpha\)/\]/\[\] blend all have velocity spans much less than that of \[\] and do not share the double-peaked distribution. The measurement is complicated by a non-linear continuum and blending of features, but generally for all emission lines except for calcium we measure maximum velocities in the region of \(\approx 3400\) km s\(^{-1}\). Nonuniform density, uneven \(^{56}\)Co heating, and chemical stratification are all possible reasons for the observed line asymmetry. An important consideration is that Ca is relatively easy to ionize (Ionization Potential = 6.11 eV) compared to Si (IP=8.51 eV) and S (IP = 10.34 eV), and that the neutral gas may be expanding more slowly. In principle if the radioactive material were evenly mixed, the heating would be proportional to \(n\) and the cooling (if optically thin) proportional to \(n^2\), so denser regions would be cooler and more neutral. The emission feature centered near 1.64 \(\mu\)m could potentially be ascribed to \[\], but the absence of the stronger \[\] line at 1.257 \(\mu\)m makes this impossible. Instead, the 1.64 \(\mu\)m line must be \[Si I\] 1.645 \(\mu\)m, an identification confirmed by the 1.6068 \(\mu\)m \[Si I\] line with the intrinsic 1:3 intensity ratio for this pair of lines. The presence of the \[\] line pair suggests that the lines at 1.08 and 1.13 \(\mu\)m could be the corresponding \[\] 1.0824 and 1.1309 \(\mu\)m lines, rather than the 1.083 and O I 1.129 \(\mu\)m lines near these wavelengths (e.g., @Jerkstrand15). The absence of the 1.317 \(\mu\)m line and the multiplet at 7774 Å, which should be 30 times stronger than the 1.129 \(\mu\)m line, rule out the identification and place an upper limit of 10% to the contribution to the 1.13 \(\mu\)m feature. The measured intensity ratio \(I\)(1.08)/\(I\)(1.13) is 4.2:1, rather than the 3.4:1 ratio given by the \[\] Einstein A values, suggesting that may indeed contribute to the 1.08 \(\mu\)m line. \[\] 1.099 may also contribute to the feature. A feature consistent with 2.058 \(\mu\)m, which should be a few times fainter than the 1.083 \(\mu\)m line, is marginally detected (Figure [\[fig:nirspec\]](#fig:nirspec){reference-type="ref" reference="fig:nirspec"}). The optical spectra indicate an upper limit to the 5876 line of \(1.3 \times 10^{-15}~\rm erg~cm^{-2}~s^{-1}\), which would indicate an upper limit of 12% contribution of He I to the 1.08 \(\mu\)m feature. This is within the measurement uncertainties of the value of 20% indicated by the difference between the observed line ratio and the intrinsic \[S I\] intensity ratio. We note that the \[\] and \[\] lines have very low excitation thresholds, so that they can be excited in fairly cool gas, while the He I and O I lines are produced by recombination and require ionized helium and oxygen. The absence of \[\] 1.03 \(\mu\)m emission while \[\] emission is strong (Figure [\[fig:NIRenlarge\]](#fig:NIRenlarge){reference-type="ref" reference="fig:NIRenlarge"}), implies either predominantly neutral sulfur or very cool gas.
## Chemical abundances {#sec:abundances}
We estimated chemical abundances of the ejecta of iPTF15eqv using the relative line strengths and luminosities of the ions identified in our optical and NIR spectra. The ratio of the Ca triplet to the \[\] lines at 7300 Å depends on temperature and density. On day 95, the ratio is 0.7, while on day 158 it has fallen to 0.1. The Ca triplet lines are produced by decay of the 4p level, which decays by way of the H and K lines at 3933, 3968 Å 94% of the time. However, for large optical depths in H and K, nearly all the H and K photons will be converted to triplet photons, which escape because of the smaller optical depth. Using atomic rates from CHIANTI version 8 and assuming that all excitations to the 4p level produce Ca triplet photons, we compute the \[Ca II\]/Ca II ratio as a function of electron density (\(n_e\)) and temperature (T). We restrict the temperature range to 3.9 \(<\) log T \(<\) 4.1. At lower temperatures a : ratio above 0.1 would imply that the \] line at \(\lambda\)`<!-- -->`{=html}6572 line would exceed the intensity of the observed 6560 Å feature (based on the \] excitation rate from @SB01). At higher temperatures Ca would be ionized to and ions such as and are expected to be strong. None of these higher ionization states are observed (see Figures [\[fig:spectra\]](#fig:spectra){reference-type="ref" reference="fig:spectra"} and [\[fig:nirspec\]](#fig:nirspec){reference-type="ref" reference="fig:nirspec"}). The allowed density range within these constraints is \(10^{5.9}\) cm\(^{-3}\) at log T = \(4.1\) to \(10^{6.7}\) cm\(^{-3}\) at log T = \(3.8\). Between day 95 and day 158, the 7300 Å flux drops by a factor of 4.6, while the Ca triplet flux drops by a factor of 20. Homologous expansion during that interval implies that the density drops by a factor of 4.6. The flux changes in both lines can be matched with a factor of 4.6 drop in density and a factor of about 1.4 drop in temperature. That is a smaller change than the factor of 2.8 from adiabatic expansion alone, indicating that radioactive heating partly compensates for adiabatic cooling. The \[\] and \[\] lines have similar excitation potentials and critical densities, thus their intensity ratio depends relatively weakly on temperature and density. In the temperature range derived above, most of the oxygen is neutral and most of the calcium is , so that the intensity ratio of the lines reflects the Ca to O abundance ratio. Since the \[\] to \[\] ratio is nearly constant at \(\approx 0.1\) during our observations, but the emissivity of \[\] is much larger than that of \[\] according to CHIANTI version 8, the O:Ca abundance ratio (by number) is 39\(\pm\)`<!-- -->`{=html}24. The substantial range is due to the range of temperature and density derived above. The relative excitation rates of \[\] and \[\] vary less with density and temperature than those of \[\] and \[\], but the uncertainty in the flux calibration of the NIR spectrum and the necessity to interpolate the \[\] fluxes between nearby observations dominate the measurement uncertainty. We also assume that 1/5 of the 1.089 \(\mu\)m emission comes from He I to match the 1.089 to 1.131 \(\mu\)m intensity ratio. We use CHIANTI version 8 to predict the emissivities. Notably, Table 3 of mislabels the collision strengths from levels 1 and 3 to levels 4 and 5, and we have corrected the corresponding data file. We find an S:Ca abundance ratio of 29\(\pm\)`<!-- -->`{=html}7. The \[\] lines suffer from a still higher calibration uncertainty and greater difficulties in removing night sky lines. Using the excitation cross section of and an Einstein A value of 0.0037, we find that the \[\] emissivity is 5 to 10 times smaller than the \[\] emissivity over the range of density and temperature that we infer. The \[\] flux is 1/3 that of the \[\] lines, suggesting the Si abundance is 1.7 to 3.3 times as large as that of S. However, since the \[Si I\] excitation rate did not include resonances or cascades, it may be an underestimate. At a distance of 30.4 Mpc, the \[\] luminosity is \(3.5 \times 10^{40}~\rm erg~s^{-1}\). With emissivities of \(2.0 \pm 0.6 \times 10^{-13}~\rm erg~s^{-1}\) per atom, that implies a calcium mass of 0.006\(\pm\)`<!-- -->`{=html}0.002 M\(_{\odot}\). The masses of oxygen, silicon, and sulfur are about \(0.09_{-0.08}^{+0.11}\), \(0.31 \pm 0.2\), and 0.14\(\pm 0.08\) M\(_{\odot}\), respectively. If we assume that the nebular emission arises in a shell given by a velocity of 3400 \(\rm km~s^{-1}\), or R=\(2.8 \times 10^{15}~\rm cm\) at day 95, and that the thickness of the shell is R/8 because of a steep (\(n \propto R^{-9}\)) density distribution, the density of ions is \(5 \times 10^6~\rm cm^{-3}\). As this lies within the electron density range derived above, it is at least plausible that calcium provides most of the free electrons in the nebular shell. There are many caveats to our estimates. Absorption may obscure interior emission, the ejecta may be clumpy, and our assumption that all lines are formed in the same volume may be incorrect. Indeed, indications that some degree of asymmetry in the ejecta not accounted for in our estimates can be seen in the clear asymmetry in the line profiles (Figure [\[fig:lineprofiles\]](#fig:lineprofiles){reference-type="ref" reference="fig:lineprofiles"}). Furthermore, there is noticeable discrepancy between the masses we obtain from fitting the light curve (section [4.1](#sec:lc){reference-type="ref" reference="sec:lc"}) and from the spectra. Our spectral analysis may not be accounting for all mass because much of the mass is at undetectable low densities and/or temperatures. This is especially true for H and/or He material that may be neutral and at temperatures that are too low for any significant excitation. Improved chemical abundances and mass estimates may be possible with a more sophisticated model that includes radiative transfer and additional assumptions about ejecta kinematics.
## Metallicity and star formation rate {#sec:metallicity}
Unlike many Ca-rich transients that are discovered in isolated locations, iPTF15eqv is within projected proximity of a fairly rich environment of hot gas and stars (Figure [\[fig:fchart\]](#fig:fchart){reference-type="ref" reference="fig:fchart"}). Using our optical spectra we estimated the metallicity of the region surrounding the explosion site of iPTF15eqv using the Markov Chain Monte Carlo method of to model the emission line spectra and measure the fluxes of the strong emission lines. Following this approach we estimate the oxygen abundance as \(\log(\rm{O} / \rm{H})+12=8.65\pm0.04\) and \(8.58\pm0.02\) using the N2 and O3N2 diagnostics, respectively, of, and as \(\log(\rm{O} / \rm{H})+12=8.82\pm0.07\) using the diagnostic of (KD02). These metallicity estimates are consistent given the characteristic offsets between the diagnostics and uncertainties, and correspond to a metallicity of \(\sim0.8~\rm{Z}_\odot\). The quoted uncertainty is not inclusive of systematic effects such as the calibration variance in the strong line metallicity diagnostics (\(\sim0.1\) dex, @KE08) or local metallicity variation in the host galaxy (as much as \(0.3\) dex, @Sanders12 and @Sanders12M31). NGC 3430 is in the imaging footprint of the SDSS, which allowed us to make use of the SDSS-calibrated Luminosity-Metallicity-Color relation (LZC) of to infer an independent estimate of galaxy metallicity from optical luminosity and color. Using the extinction corrected absolute magnitudes of the host galaxy in the SDSS DR12 cModel magnitudes, \(\rm{M}_g=-19.439\) mag and \(\rm{M}_r=-20.042\) mag, the LZC photometric metallicity for NGC 3430 is \(\log(\rm{O} / \rm{H})+12=8.68\) with a characteristic scatter of \(0.07\) dex in the N2 diagnostic calibration of the LZC relation. We also estimated the star formation rate (SFR) density of the explosion site using the \[\] and H\(\alpha\) line fluxes measured from our optical spectra. We measure an \[\] \(\lambda\)`<!-- -->`{=html}3727 brightness of \(9.7\times10^{40}~\rm{erg}~\rm{s}^{-1}~\rm{arcsec}^2\) and an H\(\alpha\) luminosity of \(1.3\times10^{41}~\rm{erg}~\rm{s}^{-1}~\rm{arcsec}^2\), with an estimated uncertainty in our spectrophotometric flux calibration of \(20\%\). Using the \[\] SFR diagnostic of, assuming a distance scale of \(0.102~\rm{kpc}~\rm{arcsec}^{-1}\), and assuming the KD02 metallicity quoted above, we find an explosion site SFR density of \((60\pm10)\times10^{-3}~\rm{M}_\odot~\rm{yr}^{-1}~\rm{kpc}^{-2}\). Using the H\(\alpha\) diagnostic of, we find a SFR density of \((80\pm20)\times10^{-3}~\rm{M}_\odot~\rm{yr}^{-1}~\rm{kpc}^{-2}\). Given the uncertainty in the flux calibration, the two estimates are similar. Both approaches indicate a vigorous SFR consistent with the morphological type of the galaxy.
## Radio synchrotron limits
The radio emission in SNe is due to synchrotron radiation. As the forward shock wave moves through surrounding circumstellar and interstellar material (CSM and ISM), a fraction of its kinetic energy gets converted into motion of the shock-accelerated particles. The relativistic electrons behind the shock are the main radiators which gyrate in the shock-amplified magnetic field and radiate synchrotron radiation. The energy distribution of these relativistic electrons is a power law, \(n_{e}(\gamma_{e}) \propto \gamma_{e}^{-p}\), with index \(p\) and the Lorentz factor \(\gamma_{e}\) of individual electrons. This form of the electron distribution is often evident from the synchrotron radio spectrum of SNe. Thus, we can utilize our deep radio observations of iPTF15eqv to probe unique information about its immediate environment. For a constant mass-loss rate and wind velocity, the density in the stellar environment falls with the distance as the inverse squared \(\rho \propto r^{-2}\). In a typical massive star, this wind-density profile may extend to several parsecs. A typical supernova shock wave traveling at the speed of a few \(\times 10^{4}\) km s\(^{-1}\) may take several decades before crossing the progenitor wind. We followed procedures outlined in, based on work from, , and, to estimate the radio emission generated by a shock wave expanding in such an environment. The deep upper limits on radio brightness evolution at 8.6 GHz constrain the progenitor wind mass-loss rate to be \(\dot{M} \la 7 \times 10^{-5}\, \rm M_{\odot}\;yr^{-1}\) assuming a constant wind velocity of \(v_{w} = 100\) km s\(^{-1}\). This is shown in Figure [\[fig:radio\]](#fig:radio){reference-type="ref" reference="fig:radio"} by the purple shaded region. We also show the expected radio light curves for the ISM density profile. Our observations constrain the ISM density to be \(\rm n \la 0.1\) cm\(^{-3}\). The width of the shaded regions represents the spread due to the \(0.01 \leq \epsilon_{B} \leq 0.1\) and \(\epsilon_{e} = 0.1\), where \(\epsilon_{B}\) and \(\epsilon_{e}\) are the fractions of kinetic energy channeled into the post-shock magnetic fields and shock accelerated relativistic electrons, respectively.
## Tidal disruption
One model of Ca-rich transients is that they are tidal disruptions of low mass WDs by IMBHs. provide observable signatures of this process at optical wavelengths. We find that our overall light curve evolution is not consistent with extrapolations of their models that generally show much flatter and nearly constant evolution in \(R\)-and \(I\)-band, and we do not have photospheric epoch spectra to compare with their synthetic spectra. However, a unique signature of tidal detonations by IMBHs is that they are accompanied by the fallback of substantial amounts of stellar debris returning to the BH that can be accreted and emit at X-ray wavelengths. Our CXO observations of iPTF15eqv can be used to constrain the likelihood of this scenario. The fallback rate can be several orders of magnitude above the Eddington limit, with the inner X-ray emitting portions of the accretion structure being potentially obscured by optically-thick material. While a jet can be produced during this hyper-accretive phase, the absence of any detected afterglow suggests that no jet was formed. The disk produced by the fallback will be able to radiate at roughly its Eddington limit during this phase, \[L_{\rm Edd} \equiv G M_{\rm h} m_{\rm p} c / \sigma_{\rm T} \simeq 10^{41}~M_{\rm h,3}\,{\rm erg~s}^{-1},\] where \(M_{\rm h,3}\) is the BH mass in units of \(10^3 \rm M_\odot\). Using the peak accretion rate and time found in simulations, the timescale that the fallback rate will remain above the Eddington limit is \[t_{\rm Edd} = 1~M_{\rm h, 3}^{-2/5} M_{\rm WD,0.6}^{1/5} R_{\rm WD,0.6}^{3/5}~\rm yr,\] where \(M_{\rm WD,0.6}\) and \(R_{\rm WD,0.6}\) are the mass and radius of an \(0.6 \rm M_\odot\) WD. make the point that viscous effects may act to spread the accretion of matter over somewhat longer timescales, prolonging the time that the accretion rate remains above Eddington up to a factor of \(\sim 10\). After this phase, the accretion luminosity should decay approximately as \((t/t_{\rm peak})^{-5/3}\). observed SN 2012hn 533 days after the explosion to look for accretion emission during the decay phase. No emission was detected, and from an X-ray upper limit on the luminosity of \(L_{\rm 0.5-8kev} < 4.4 \times 10^{38}\; {\rm erg\;s}^{-1}\) they concluded that SN 2012hn was unlikely to be a tidal detonation by an IMBH unless the BH is at the low end of the IMBH mass distribution. Our upper-bound limit on the luminosity (\(<5.3\times 10^{38}\,\rm{erg\,s^{-1}}\)) of iPTF15eqv is comparable to that of SN 2012hn, but since our observation of iPTF15eqv was closer to the time of explosion than SN 2012hn (114--164 days versus 533 days) we can place even stricter constraints on the IMBH mass. Given that we would expect the accretion disk to be radiating at close to its Eddington limit for nearly a year and decline rapidly afterwards, and that even the 10% accretion efficiency expected in the low-hard state would have been detected, we can restrict the BH mass to the point that only a stellar mass BH (\(\lesssim 100 M_\odot\)) could yield \(t_{\rm Edd} < 164~{\rm days}\). Our constraint on the BH mass falls into the regime of a "micro" tidal disruption event, a scenario that would only be likely in a dense stellar cluster where close encounters between stars are common. In these models a sufficiently massive WD is tidally disrupted by its BH companion as the binary orbit shrinks due to unstable mass transfer. The WD debris is sheared into an accretion disk with an initial size that is comparable to that of the initial binary at the time of Roche-lobe overflow.
# Discussion {#sec:discussion}
iPTF15eqv exhibits a unique combination of properties that bridge those observed in Ca-rich transients and SNe Ib/c. On the one hand, the late-time spectra of iPTF15eqv are dominated by strong calcium emission lines, which is a defining characteristic of Ca-rich transients. We illustrate this similarity in Figure [\[fig:ca-rich-compare\]](#fig:ca-rich-compare){reference-type="ref" reference="fig:ca-rich-compare"}, where we show the day 199 spectrum of iPTF15eqv and a sample of Ca-rich transients from the literature. On the other hand, iPTF15eqv differs from other Ca-rich transients in terms of its light curve evolution. iPTF15eqv was a slower evolving and potentially much more luminous object than Ca-rich transients and exhibited a light curve that resembles SNe Ib/c (Figure [\[fig:ca-rich-lc\]](#fig:ca-rich-lc){reference-type="ref" reference="fig:ca-rich-lc"}). Below we explore how to place iPTF15eqv within the Ca-rich transient classification and seek to understand whether a single progenitor system framework can unite their diverse properties.
## \[\]/\[\]
For Ca-rich transients having optical spectra publicly accessible through the Weizmann Interactive Supernova Data Repository (WISEREP; @YG12) and the Open Supernova Catalog (OSC; @Guillochon17), we measured the \[\]/\[\] relative flux ratio (for cases where both lines could be observed) to determine the range of ratios encountered thus far in the literature. Multiple epochs were measured when possible. The results are listed in Table 4 and plotted in Figure [\[fig:CaIIvOI\]](#fig:CaIIvOI){reference-type="ref" reference="fig:CaIIvOI"}. For comparison, we also measured this ratio for a number of H-rich (Type II) and H-poor (Type IIb/Ib/Ic) SNe. At all epochs where both lines are visible and the conditions can be reasonably assessed to be nebular, all Ca-rich transients have \[\]/\[\] \(> 2\) and all Type Ib/c are \(<2\). iPTF15eqv exhibits the strongest \[\]/\[\] emission ratio (\(\approx 10\)) in our sample. SNe II typically have higher \[\]/\[\] ratios between \(\approx 2-4\). However, the phase at which this ratio is observed in Type II supernovae (\(t > 200\) d) is much later than Ca-rich transients (\(50 < t < 150\) d). Notably, the ratio of \[\]/\[\] for many Type Ib/c supernovae can decrease with time. For example, a 30% change in the ratio was noted in SN 2004gq between day 258 and 381. For this reason, raised the issue that the high \[\]/\[\] ratio seen in some Ca-rich transients at relatively early phases may in part be a phase-dependent effect rather than reflecting an intrinsic abundance pattern. In the case of iPTF15eqv, the \[\]/\[\] ratio exhibited modest change between the the epochs where both features were well observed (i.e., days 95-199; Figure [\[fig:spectra\]](#fig:spectra){reference-type="ref" reference="fig:spectra"}). This lack of change is consistent with observations of SN Ib/c at very evolved epochs where change is routinely negligible, and in line with the fact that the critical densities for \[\] and \[\] are comparable. Although we have not explored epochs \(t > 300\) d, a decrease in \[\]/\[\] ratio alone appears to be an unlikely explanation for why Ca-rich transients have strong calcium line emission.
## Spectral fingerprinting {#sec:late-time}
Optical and NIR spectroscopy obtained many months to years after maximum light is a powerful probe of the kinematic and chemical properties of the ejecta closest to the explosion center. It can be used to discriminate between Type Ia versus core-collapse classifications, and even between sub-classifications, and thus provides a unique means of spectrally fingerprinting a supernova's progenitor system. Few Ca-rich transients have had high-quality, multi-epoch data at late times. The close proximity and relatively slow decline rate of iPTF15eqv enabled spectroscopic observations of high signal-to-noise ratio at epochs \(>\) 200 days after explosion, which is among the latest epochs ever observed for a Ca-rich transient. The high quality of the spectra reveals faint features that would be missed in lower signal-to-noise data. We compared the day 199 spectrum of iPTF15eqv to prototypical Type I and II SNe (Figure [\[fig:haha\]](#fig:haha){reference-type="ref" reference="fig:haha"}). Spectra were arbitrarily scaled to compare and contrast the relative strengths of specific emission features between nebular epoch spectra of iPTF15eqv and other supernovae. Clearly, iPTF15eqv and the normal, well-observed Type Ia SN 2011fe have widely different late-time emissions. Type Ia late-time spectra are dominated by a complicated blend of forbidden and permitted iron lines between 4000 and 5500 Å, many of which are associated with Fe-peak elements. Peaks centered around 7000-7600 Å associated with iron and nickel are also regularly observed. None of these features are conspicuous in iPTF15eqv. In sharp contrast, iPTF15eqv shares many spectral signatures of core-collapse SNe. The late-time spectrum of the Type IIb SN 2011dh, exhibits \[\], \[\], a blend of +, a "plateau" of emission spanning 3800-5600 Å associated with Fe-peak elements, and a broad H\(\alpha\) emission feature redward of \[\] (that may have some contribution from \[\]), which overlap with iPTF15eqv. The noticeable difference is in the strength of the Fe-peak plateau and the \[\]/\[\] ratio, as well as the \] emission which is prominent in many SNe IIb but not so in iPTF15eqv. A surprising similarity is also found between iPTF15eqv and SN 1998bw, which is a broad-lined Type Ic that was associated with the long-duration \(\gamma\)-ray burst GRB 980425. In this case, the relative strength of the Fe-peak plateau is comparable between the two, but the \[\]/\[\] is much smaller. Considerable overlap between the late-time spectra of the Type IIP SN 2004et and iPTF15eqv is also evident. However, the H\(\alpha\) line dominates the spectrum of SN 2004et (\[Ca II\]/H\(\alpha\) = 0.6), and is far stronger relative to that observed in iPTF15eqv (\[Ca II\]/H\(\alpha\) = 30; but see also section [5.3](#sec:hydrogen){reference-type="ref" reference="sec:hydrogen"} for discussion of other emission lines that may contribute to this wavelength region). Notably, the \[\] 7291, 7324 emission line profile of iPTF15eqv is unlike any other Ca-rich transient (Figures [\[fig:lineprofiles\]](#fig:lineprofiles){reference-type="ref" reference="fig:lineprofiles"} and [\[fig:ca-rich-compare\]](#fig:ca-rich-compare){reference-type="ref" reference="fig:ca-rich-compare"}), and rare among SNe Ib/c. To our knowledge, the only instance is in late-time spectra of the Type Ib SN 2007uy, though in that case the velocity span was much smaller (\(\la 5000\) km s\(^{-1}\); @Milisavljevic10 [@Roy13; @Modjaz14]). Double-peaked profiles have been observed in the \[\] \(\lambda\lambda\)`<!-- -->`{=html}6300, 6364 line profiles of SNe Ib/c and interpreted to represent ejecta asymmetry, possibly in the form of an O-rich torus, that may have been shaped by a jetted explosion. However, many of these cases are actually due to the doublet nature of \[\] in combination with significant density enhancements associated with ejecta clumping.
## Hydrogen in Ca-rich transients {#sec:hydrogen}
The strength of the 6560 Å feature relative to \[\] observed in iPTF15eqv is broadly similar to that observed in SN 2011dh and other Type IIb supernovae. We identify this feature to be primarily H\(\alpha\) in iPTF15eqv, however, we have noted that \] 6572 and \[\] 6548, 6583 may also contribute emission in this region. In Figure [\[fig:sn2011dh\]](#fig:sn2011dh){reference-type="ref" reference="fig:sn2011dh"} we show the day 199 spectrum of iPTF15eqv and a day 292 spectrum of SN 2011dh. The emission in iPTF15eqv is more narrowly peaked, which reflects differences in how the presumably H-rich ejecta are distributed in the two objects. A parabolic distribution is anticipated from a spherical geometry. A flat-topped distribution as observed in SN 2011dh is reflective of a shell-like geometry. iPTF15eqv is not unique in potentially exhibiting H\(\alpha\) emission among Ca-rich transients. The subluminous and peculiar PTF09dav was reported to exhibit H\(\alpha\) with a velocity width of \(\sim 1000\) km s\(^{-1}\) in nebular spectra presented in. The H\(\alpha\) emission was interpreted as circumburst material being photoionized from interaction between the explosion and the progenitor star's wind. The scenario for PTF09dav is different from what we propose for iPTF15eqv, where H\(\alpha\) exhibits higher velocities that are more easily understood as being associated with explosive ejecta. The possibility of a contribution from \] 6572 was not discussed within the context of PTF09dav. In Figure [\[fig:sn2011dh\]](#fig:sn2011dh){reference-type="ref" reference="fig:sn2011dh"}, we also show the Ca-rich transient PTF11kmb enlarged around the same region. Intriguingly, PTF11kmb exhibits excess emission redward of \[\] that resembles the H\(\alpha\) distribution of SN 2011dh.
## High velocity kicks?
To explain how Ca-rich transients explode in isolated regions devoid of massive star populations, proposed long-lived progenitor systems that have been "kicked" to high velocities. Similarly, argued that the progenitors of Ca-rich transients are double WD systems that have been ejected from their host galaxies after interacting with a super-massive BH. analyze the nebular spectra of 13 Ca-rich transients to look for evidence in support of this hypothesis in the \[\] lines and find a correlation between projected distance from the host galaxy and bulk Doppler shift velocity. Specifically, SNe with small projected offsets have large line-of-sight velocity shifts as determined by nebular lines, while those with large projected offsets have no significant velocity shifts. We considered the implications of on the unusual double-peaked \[\] profile of iPTF15eqv and found an important caveat in this interpretation. Although the \[\] line profile exhibits clear asymmetry potentially associated with a bulk Doppler shift, the profiles of all other ions of iPTF15eqv are centered near zero rest velocity (Figure [\[fig:lineprofiles\]](#fig:lineprofiles){reference-type="ref" reference="fig:lineprofiles"}). This characteristic of iPTF15eqv is not unique, in that the \[\] emission line profile of SN explosions at nebular epochs can often have an apparent blueshift attributable to internal absorption of inner ejecta. If the ejecta are not completely optically thin, then emission from the rear of the expanding debris cloud is preferentially blocked. This effect, which can affect some ions more than others, gives the overall impression of being blueshifted. We illustrate this phenomenon of apparent blueshifts in Figure [\[fig:shift\]](#fig:shift){reference-type="ref" reference="fig:shift"}, where we show the emission line profiles of \[\] and \[\] for two Type Ib/c supernovae and two Ca-rich events, in addition to that of iPTF15eqv. \[\] emission has a large range distributions that sometimes reflects the same distribution as \[\] (SN 2013ge and PTF11kmb), but other times can show an apparent bulk blueshift (SN 2002ap and SN 2012hn). interpret the \[\] profile of SN 2012hn to be associated with a systemic velocity of \(-1730 \pm 70\) km s\(^{-1}\). However, this view is not easily reconciled with the emission line profile of \[\] that is distributed about zero velocity. Thus, we do not find strong evidence supporting a high systemic bulk velocity in iPTF15eqv and conclude that other Ca-rich transients may be similarly affected by optical depth effects. Interestingly, the blueshifted \[\] emission of iPTF15eqv and SN 2012hn trace similar distributions. SN 2012hn, however, does not have a prominent redshifted emission peak like that observed in iPTF15eqv. This could be consistent with a shared torus geometry, if in the case of SN 2012hn its rear hemisphere has been obscured. This speculative scenario suggests that an observer's line of sight may be an important consideration in interpreting spectra of Ca-rich transients.
## Metallicities of Ca-rich transient host galaxies
Although indirect lines of evidence from host galaxy samples and explosion site locations point toward low mass and low metallicity progenitor stars, no systematic observational survey of the host galaxy metallicities or explosion site star formation rates of Ca-rich SNe has been conducted. We examined 5 Ca-rich transients (SN 2000ds, 2001co, 2003dg, 2005E, and PTF11bij) that have host galaxies in the SDSS imaging footprint and for which we could apply the LZC-relation of to estimate galaxy metallicity from optical luminosity and color. The foreground-extinction-corrected absolute SDSS magnitudes for each host galaxy and the host galaxy global metallicities derived from the LZC relation on the N2 scale of are displayed in Table 5. Host galaxies for four of the five Ca-rich transients (SN 2000ds, 2001co, 2003dg, and PTF11bij) have \(\log(\rm{O} / \rm{H})+12\sim8.7\), which is consistent with the metallicity reported here for iPTF15eqv. SN 2005E is the sole outlier with a substantially lower global metallicity (\(\log(\rm{O} / \rm{H})+12=8.4\)). Thus, we conclude that Ca-rich transients often occur in metal-rich galaxies. Notably, this finding is not at odds with a low-mass star progenitor scenario. The galaxy mass-normalized rate of Type Ia supernovae is actually highest, marginally, in Type Sc host galaxies among all major morphological types (@Li11Hosts; see also @Lyman13).
## Ca abundance and enrichment
The potentially large calcium abundance of Ca-rich transients, coupled with their large galactocentric distances, make them intriguing candidate drivers of chemical evolution in galaxies and the intergalactic medium. Recent works suggest that combinations of Type Ia and core-collapse nucleosynthesis models under-produce the amount of calcium required by X-ray observations of the intra-cluster medium, and that Ca-rich transients may play a role in addressing this problem. Ca-rich transients have also been invoked to explain metal-poor stars with an extremely large over-abundance of Ca, such as SDSS J234723.64+010833.4 discovered in the outer halo of our galaxy. We derived O and Ca masses of \(0.090_{-0.08}^{+0.11}\) and 0.006\(\pm\)`<!-- -->`{=html}0.002 M\(_{\odot}\), respectively, for iPTF15eqv from our nebular spectra. In contrast, estimates of chemical abundances of SN 2005E by were 0.037 and 0.135 M\(_{\odot}\) for O and Ca, respectively. Taking into account that the \[\] and \[\] emission lines of iPTF15eqv are \(\sim 2.5\) and \(\sim 20\) times more luminous than SN 2005E, the O abundance estimates are broadly consistent, whereas the Ca abundances are not. Some of the inconsistency may be attributable to different model assumptions, but additional discrepancies may arise from the different natures of the explosions. Consequently, the Ca abundance yield of Ca-rich transients may vary considerably across individual events. Regardless of the exact nucleosynthetic yields, both analyses point to a relatively large Ca/O abundance ratio, which is generally understood to be sensitive to the progenitor mass in core-collapse SNe. Models of supernova nucleosynthesis have routinely shown that larger masses of Ca are made for lower mass progenitors. For example, theoretical models predict that stars having main-sequence masses of 13 M\(_{\odot}\) and 18 M\(_{\odot}\) produce Ca/O ratios (by mass) of 0.025 and 0.005, respectively. Keeping with this trend, in order to produce a relatively large Ca/O ratio of \(\approx 0.07\) observed in iPTF15eqv, the mass of its progenitor star would have been \(< 13\) M\(_{\odot}\) and thus approach the threshold for core collapse (\(8\pm1\) M\(_{\odot}\); @Smartt09).
## Electron capture supernova?
Stars in the mass range \(8-12\) M\(_{\odot}\) can exhibit structural peculiarities during their evolution that considerably affect the supernova explosion dynamics if they undergo core collapse. Electron captures on \(^{24}\)Mg and \(^{20}\)Ne in a degenerate O-Ne-Mg core of mass ∼\(\sim 1.37\) M\(_{\odot}\) can drive the core towards collapse in what is referred to as an "electron-capture supernova" (ECSN; @Nomoto84). This progenitor scenario was invoked to explain the Ca-rich SN 2005cz. The winds of such stars are not strong enough to remove hydrogen envelopes, which led to suggest that binary interaction via Roche-lobe overflow or common envelope ejection could explain why no conspicuous hydrogen was present in their optical spectra of SN 2005cz. A similar progenitor scenario could apply to iPTF15eqv. Generally, numerical simulations of ECSNe explode with kinetic energies of \(\sim 10^{50}\) erg and synthesize \(\sim 10^{-3}\) M\(_{\odot}\) of radioactive \(^{56}\)Ni (e.g., @Kitaura06). These values are each approximately an order of magnitude less than those we infer for iPTF15eqv (section [4.1](#sec:lc){reference-type="ref" reference="sec:lc"}). However, there are many uncertainties with the explosion processes of ECSNe (see, e.g., @Jones13 [@Woosley15]). Perhaps the strongest signature of an ECSN is that it likely leaves behind a very special abundance pattern in its ejecta (see the recent review by @Mueller16), which may be abundant in isotopes of calcium. investigated the expected rates and bolometric light curve properties of stripped-envelope ECSNe using stellar models from the Binary Population and Spectral Synthesis (BPASS) code[^7]. We compared the light curve of iPTF15eqv to the Type IIb, Z=0.020 models and were unable to find a strict match. However, the models are fixed at \(2.5 \times 10^{-3}\) M\(_{\odot}\) of \(^{56}\)Ni, which mainly changes the peak luminosity. By scaling the luminosity of a model having a total ejecta mass of 1.5 M\(_{\odot}\) and hydrogen envelope mass of 0.35 M\(_{\odot}\), reasonable agreement is possible if M\(_{\rm Ni}\) is allowed to range between \(0.045-0.055\) M\(_{\odot}\) (Figure [\[fig:me16\]](#fig:me16){reference-type="ref" reference="fig:me16"}). The range in mass of our fit is due to uncertainty in the explosion date. These results are consistent with our independent derivation of M\(_{\rm Ni}\) in section [4.1](#sec:lc){reference-type="ref" reference="sec:lc"}. Notably, binary systems of stars in the mass range that develop to ECSNe can significantly extend the delay time distribution of core-collapse SNe. For single stars the maximum lifetime is approximately \(\sim 50\) Myr. However, the majority of binary progenitors in the models have lifetimes of 30 to 100 Myr, and the oldest lifetime is 200 Myr. A population synthesis study conducted by to compute the delay time distribution of core collapse SNe also found that a significant fraction can occur 50-200 Myr after birth, well after all massive single stars have already exploded. These late events originate mostly from binary systems with both stars of mass 4-8 M\(_{\odot}\).
## Ultra-stripped supernova?
Ultra-stripped SNe are core-collapse SNe whose ejecta masses are only \(\sim 0.1~M_\odot\). The small ejecta masses result from the extreme stripping of their progenitors due to the existence of compact companion stars. The ultra-stripped SNe are found to explode successfully with the neutrino-driven mechanism, with explosion energies of \(\sim 10^{50}~\mathrm{erg}\) and \(^{56}\)Ni masses of \(\sim 0.01~M_\odot\). found that light curves and spectra of some Ca-rich transients match well to those predicted by ultra-stripped SNe, suggesting some Ca-rich transients may be ultra-stripped SNe. Ultra-stripped SN light curve models currently available tend to decline faster than that of iPTF15eqv. However, the current light curve models are only available for progenitors with ejecta masses below \(0.2~M_\odot\). There exist ultra-stripped SNe with larger ejecta masses and they can have slower decline rates, possibly matching iPTF15eqv. One problem applying the ultra-stripped model to iPTF15eqv is that iPTF15eqv likely exhibits hydrogen in our late-time spectra (sections [4.2](#sec:identify){reference-type="ref" reference="sec:identify"} and [5.3](#sec:hydrogen){reference-type="ref" reference="sec:hydrogen"}; Figure [\[fig:sn2011dh\]](#fig:sn2011dh){reference-type="ref" reference="fig:sn2011dh"}). Ultra-stripped SN progenitors lack hydrogen because of the extreme stripping, and their companions are compact stars that do not have hydrogen. Therefore, no hydrogen is expected to exist in the progenitor system. We consider an ultra-stripped SN origin unlikely for iPTF15eqv, although further work on this progenitor scenario is needed.
## Mass loss environment
Mass loss is a key process in stellar evolution, and thus is a powerful probe of the progenitor systems of SNe. Observations of SNe at radio and X-ray wavelengths can reconstruct the mass loss history of its progenitor system as the SN blast wave passes through and interacts with surrounding material. To date no Type Ia supernova has been detected at X-ray or radio wavelengths, suggesting a relatively clean environment with densities below \(n < 3\) cm\(^{-3}\) at \(R \sim 10^{16}\) cm and progenitor system mass loss rates of \(\dot{M} < 10^{-9}\, \frac{M_{\odot}\;\rm yr^{-1}}{100\; \rm km\;s^{-1}}\). In contrast, core-collapse SNe exhibit a wide range of X-ray and radio properties that reflect the diversity of their environments that have been sculpted by the progenitor system's mass loss. examined radio upper limits for Ca-rich transients in the context of Type Ia thermonuclear WD explosions. Assuming a total ejecta mass of 0.3 M\(_{\odot}\) and kinetic energy of \(4 \times 10^{50}\) erg, they constrain \(\dot{M} < 10^{-5}\, \rm M_{\odot} yr^{-1}\) for \(v_{w} = 100\) km s\(^{-1}\) for most objects of their sample. This allows them to rule out the presence of tidal tails, in the case where stripping of the He white dwarf occurred a few years before SN explosion, as well as the presence of strong accretion-powered outflows or winds, as might be expected if a disrupted white dwarf is accreted. In light of the overlap in properties between iPTF15eqv and core-collapse SNe (section [5.2](#sec:late-time){reference-type="ref" reference="sec:late-time"}), we compared our VLA upper limits on radio emission from iPTF15eqv to the radio light curves of SNe IIb and Ib/c (Figure [\[fig:radio-compare\]](#fig:radio-compare){reference-type="ref" reference="fig:radio-compare"}). Our observations rule out environments similar to those observed around the Type IIb SN 1993J and SN 2011dh, which exhibited radio light curves that are more luminous than iPTF15eqv and were estimated to have mass loss rates ranging between \(\dot{M} \approx 2.4 \times 10^{-5}\) M\(_{\odot}\) yr\(^{-1}\) (\(v_w = 10\) km s\(^{-1}\)) and \(\dot{M} \approx 6\times 10^{-5}\) M\(_{\odot}\) yr\(^{-1}\) (\(v_w = 1000\) km s\(^{-1}\)), respectively. The environments of other Type IIb such as SN 2008ax with \(\dot{M} = (9 \pm 3) \times 10^{-6}\) M\(_{\odot}\) yr\(^{-1}\) (for \(v_w = 10\) km s\(^{-1}\)) are not ruled out. The difference in mass loss rates of SNe IIb may reflect the size and evolutionary status of the progenitor star. Many SNe Ib/c such as SN 2013ge and SN 2002ap with relatively clean environments are not ruled out, but late-interacting SNe Ib/c such as SN 2014C are ruled out. Our deep non-detections up to \(\sim\) 400 days imply a low-density environment (\(n \la 0.1\) cm\(^{-3}\)) for the progenitor of iPTF15eqv out to \(R \sim 10^{17}\) cm. Using a typical forward shock velocity of \(\approx 3 \times 10^4\) km s\(^{-1}\) for core-collapse SNe, we estimate that the progenitor system of iPTF15eqv did not undergo any enhanced mass loss or eruptive activity within \(\sim 300-1000\) (\(v_w/100\) km s\(^{-1}\)) yr of the SN.
# Conclusions {#sec:conclusions}
We have presented multi-wavelength observations and analysis of the supernova explosion iPTF15eqv in NGC 3430. We find that iPTF15eqv has a combination of properties that bridge those observed in Ca-rich transients and SNe Ib/c. Perhaps most revealing about iPTF15eqv is its large \[\]/\[\] \(\approx 10\) emission line ratio, which is among the highest encountered among Ca-rich events, and clear Type Ib/c and Type II spectroscopic signatures (Figure [\[fig:haha\]](#fig:haha){reference-type="ref" reference="fig:haha"}) seen in our optical and NIR data \(\ga 200\) days post-explosion. This discovery is significant because discussions of Type Ia SNe often include Ca-rich transients (e.g., @Parrent14 [@Chomiuk16]), and our work establishes that an unknown percentage of SNe that have been considered to be Ca-rich transients may be associated with massive star explosions. Our data allow us to place many constraints on the chemical abundance yields of the explosion and the nature of the progenitor system of iPTF15eqv:
1. We estimate M\(_{\rm ej} \approx 2-4\) M\(_{\odot}\), M\(_{Ni}\) \(\approx 0.04-0.07\) M\(_{\odot}\), and E\(_{k}\) \(\approx (0.8-2) \times 10^{51}\) erg, assuming an explosion date sometime between 2015 July 24 and September 12.
2. We derive an O:Ca abundance ratio (by number) of 39\(\pm\)`<!-- -->`{=html}24, S:Ca abundance ratio of 29\(\pm\)`<!-- -->`{=html}7, and Si:S abundance ratio of \(\sim 1\).
3. Upper limits on possible radio emission from iPTF15eqv made with the VLA suggest a low density environment (\(n \la\) 0.1 cm\(^{-3}\)) within a radius of \(\sim 10^{17}\) cm of the explosion site.
4. We do not observe X-ray emission that would be associated with accretion onto a BH to an upper limit of \(5.3\times 10^{38}\,\rm{erg\,s^{-1}}\), which constrains the possibility of any BH to a mass \(\lesssim 100 M_\odot\).
5. The leading candidate progenitor system is an H-poor star of mass \(\la 10\) M\(_{\odot}\) that interacted with a nearby companion star. The chemical abundances and ejecta kinematics inferred from our late-time spectra suggest that the explosion may have been an ECSN.
iPTF15eqv forces a sober reconsideration of its "Ca-rich transient" classification. On the one hand, without early epoch photometry and spectra to constrain its light curve and specific SN classification, iPTF15eqv does not strictly fit the criteria of a "SN 2005E-like" or "Ca-rich gap transient". However, iPTF15eqv does exhibit strong calcium line emission seen in other SNe having incomplete spectral and photometric characterization (including 2003H, 2003dr, 2003dg, 2001co, 2000ds) that have been investigated as Ca-rich transients. It is noteworthy that objects considered to be Ca-rich transients thus far all share spectral evolution that truly distinguishes them from other SNe (see Figures [\[fig:ca-rich-compare\]](#fig:ca-rich-compare){reference-type="ref" reference="fig:ca-rich-compare"} and [\[fig:CaIIvOI\]](#fig:CaIIvOI){reference-type="ref" reference="fig:CaIIvOI"}), which is an indication that the observational classification is meaningful. As data improve in quality and wavelength span (particularly in the NIR; see section [4.2](#sec:identify){reference-type="ref" reference="sec:identify"}) the ability to define a further subdivision in the class may be achievable, as was done in the Type I classification of that was later split to Type Ia (thermonuclear explosions of WDs) and Type Ib/c (massive star core collapses). However, reconciling how two radically different progenitor channels can exhibit similar emission properties at nebular epochs when emission is from interior ejecta will be challenging (Figure [\[fig:haha\]](#fig:haha){reference-type="ref" reference="fig:haha"}). The relatively long delay-time distribution and distinguishable abundance patterns of ECSNe from binary systems make them an attractive progenitor system for at least some Ca-rich transients. It is important to note, however, that alternative models not immediately applicable to iPTF15eqv, including "ultra-stripped" SNe and violent disruptions of WDs, successfully reproduce other observed properties of Ca-rich transients. Additional theoretical work that can provide distinct signatures of the various progenitor scenarios is needed. In the case of iPTF15eqv, high-resolution imaging obtained after the SN has faded can map out the surrounding stellar environment and be used to investigate whether a massive star association is nearby. More generally, NIR spectra at nebular epochs proved to be critical for our line identifications and diagnostics, and it is hoped that future Ca-rich transients can be observed and analyzed in this way. Perhaps the most pertinent issue for Ca-rich transients is to quantify their calcium abundances more accurately. Given that present estimates of their yields vary across orders of magnitude, more stringent constraints on their progenitor systems and improved models of their late-time emissions are needed if a precise understanding of the chemical enrichment of galaxies and the intracluster medium is to be achieved. We thank the referee H. Perets for a critical reading of our manuscript and for providing many suggestions that improved its content and presentation. Observations reported here were obtained at the MMT Observatory, a joint facility of the Smithsonian Institution and the University of Arizona. Based on observations made by the *Chandra X-ray Observatory* under program GO 16500534, PI Margutti, observations ID 16821. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. This paper made use of the Weizmann interactive supernova data repository (WISEREP)-http://wiserep.weizmann.ac.il-and the Open Supernova Catalog-http://sne.space. K. Kawabata kindly provided data of SN 2005cz. T.J.M. is supported by the Grant-in-Aid for Research Activity Start-up of the Japan Society for the Promotion of Science (16H07413). R.M. gratefully acknowledges support from the Research Corporation for Science Advancement. The authors acknowledge the Telescope Data Center of the Smithsonian Astrophysical Observatory for performing MMIRS data reduction. I.C. acknowledges partial support by the Russian--French PICS International Laboratory program (no. 6590) co-funded by the RFBR (project 15-52-15050), by the RFBR project 15-32-21062 and the presidential grant MD-7355.2015.2. This work is based in part on observations obtained at the MDM Observatory, operated by Dartmouth College, Columbia University, Ohio State University, Ohio University, and the University of Michigan.
[^1]: http://www.cbat.eps.harvard.edu/unconf/followups/\
J10520833+3256394.html
[^2]: Y. Cao also kindly provided the original classification spectrum for us to inspect.
[^3]: http://www.cfa.harvard.edu/mmti/wfs.html
[^4]: Specifications of the KAF 1001E detector can be found at http://www.onsemi.com/pub_link/Collateral/KAF-1001-D.PDF
[^5]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
[^6]: PYRAF is a product of the Space Telescope Science Institute, which is operated by AURA for NASA.
[^7]: http://bpass.auckland.ac.nz | {'timestamp': '2017-10-16T02:01:01', 'yymm': '1706', 'arxiv_id': '1706.01887', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01887'} |
# Introduction
The strong coupling is one of the fundamental parameters of the standard model of particle physics thus its precise determination is mandatory. Nowadays, event shapes and jet rates measured in three-jet formation in electron-positron annihilation are still among the most precise tools used for accurate extractions of from data. In these analyses, the measurement of involves fitting theoretical predictions for a given observable and collider energy to observations. Hence theoretical input is essential and the goodness of the fitting procedure relies heavily on the quality of the theoretical predictions used. In high-energy particle physics we may approach the quality of theoretical predictions from two perspectives. First, the general framework for performing calculations is perturbation theory, where in case of QCD the perturbative parameter is the strong coupling, . Due to the complexity of calculations only the first few terms of the perturbative series can be evaluated and this truncation introduces a theoretical uncertainty manifested by the dependence of the predictions on non-physical parameters such as the renormalization and factorization scales. Thus one way of increasing the theoretical precision of the calculations is by including exact higher order corrections in perturbation theory. Second, the actual calculations of physical observables involve numerical integrations over the physical phase space and this introduces a statistical uncertainty in the predictions. Hence formal higher order precision must be supplemented by good numerical accuracy to obtain results that are useful for experimental needs. For three-jet production in electron-positron annihilation, all matrix elements necessary for the computation of next-to-next-to-leading order (NNLO) corrections have been known in the literature for some time and indeed NNLO corrections to several event shapes and jet rates have been evaluated . Hence this process is not only interesting from a phenomenological point of view, but it also provides an ideal testing ground for new computational methods at this order in perturbation theory. In this contribution we summarize a completely local subtraction method, called , for computing QCD corrections to jet cross sections at NNLO accuracy and present the application of our framework to three-jet production in electron-positron annihilation. Our method demonstrates excellent numerical stability and accuracy for the considered observables.
# The method
In perturbative QCD the expansion of a jet cross section defined by some physical quantity \(J\) can be formally written up to NNLO accuracy as \[\begin{aligned}
\sigma[J] &= \sigma^{\rm LO}[J] + \sigma^{\rm NLO}[J] + \sigma^{\rm NNLO}[J] + \dots \. \end{aligned}\] Focusing on the production of \(m\) jets from a colorless initial state, the leading order (LO) cross section is simply given by integrating the fully differential Born cross section for the production of \(m\) partons over the \(m\)-parton phase space defined by the observable \(J\), \[\begin{aligned}
\sigma^{\rm LO}[J] &= \int_m{\rm d}\sigma^{\rm B}_{m}J_{m} \. \end{aligned}\] The next-to-leading order (NLO) correction can be written as the sum of two terms, \[\begin{aligned}
\sigma^{\rm NLO}[J] &= \int_{m+1}{\rm d}\sigma^{\rm R}_{m+1}J_{m+1} + \int_{m}{\rm d}\sigma^{\rm V}_{m}J_{m} \,, \end{aligned}\] and is finite for any infrared-safe observable by the general theorem of Kinoshita, Lee and Nauenberg. Although the KLN theorem guarantees the finiteness for the sum of the real emission (\(\sigma^{\rm R}_{m+1}\)) and virtual (\(\sigma^{\rm V}_{m}\)) corrections, it does not say anything about the contributions separately which are indeed infinite in four spacetime dimensions. Using conventional dimensional regularization in \(d = 4-2\epsilon\) dimensions to regularize the two pieces, the singularities become poles in \(\epsilon\), which nevertheless cancel between the two contributions in the final result. However, this cancellation is not manifest. On one hand the singularities in the real emission part have a kinematic origin: they are due to divergent phase space integrals when one final state parton becomes unresolved. On the other hand, the \(\epsilon\)-poles in the virtual correction arise from the integration over the loop momentum. In general the squared matrix elements and observables in QCD are much too complicated to perform an analytic calculation in \(d\) dimensions, and our aim is to carry out the computations in four dimensions using Monte Carlo integration techniques. To do so, the real and virtual emission contributions have to be made finite separately which we achieve by local subtractions. In this method, an approximate differential cross section, \({\rm d}\sigma^{\rm R,A_1}_{m+1}\), is subtracted from the real emission contribution. This approximate cross section is constructed carefully to have the same kinematic singularity structure (in \(d\) dimensions) as the real emission cross section. Thus, the difference is free of non-integrable kinematic singularities and the phase space integral can be evaluated in four dimensions using standard Monte Carlo techniques. The poles appearing in the virtual contribution are then removed by adding back the approximate cross section after integrating over the momentum and summing over the quantum numbers (color, flavor) of the unresolved particle (these operations are collectively denoted by \(\int_1\)). Then the NLO correction takes the form \[\begin{aligned}
\sigma^{\rm NLO}[J] =& \int_{m+1}\left[{\rm d}\sigma^{\rm R}_{m+1}J_{m+1}-{\rm d}\sigma^{\rm R,A_1}_{m+1}J_{m}\right]_{d=4} \notag\\ &+ \int_{m}\left[{\rm d}\sigma^{\rm V}_{m}J_{m} + \int_1{\rm d}\sigma^{\rm R,A_1}_{m+1}J_{m}\right]_{d=4}\,, \end{aligned}\] where now both contributions are finite as discussed. Several explicit constructions are available for the approximate cross section \({\rm d}\sigma^{\rm R,A_1}_{m+1}\) in the literature. The NNLO correction is composed of three different contributions, \[\begin{aligned}
\sigma^{\rm NNLO}[J] = \int_{m+2}{\rm d}\sigma_{m+2}^{\rm RR}J_{m+2} + \int_{m+1}{\rm d}\sigma_{m+1}^{\rm RV}J_{m+1} + \int_{m}{\rm d}\sigma_{m}^{\rm VV}J_{m} \. \end{aligned}\] The first term is the double real (RR) piece which involves tree level squared matrix elements with \(m+2\)-parton kinematics and develops singularities when one or two partons become unresolved. The second term is the real-virtual one (RV) and contains the interference of one-loop and tree level matrix elements with \(m+1\)-parton kinematics. This contribution develops both kinematic singularities when a parton becomes unresolved and also contains explicit \(\epsilon\)-poles coming from one-loop amplitudes. Finally, the third term is the double virtual (VV) contribution, which includes the interference of the \(m\)-parton two-loop and tree level matrix elements as well as the square of the \(m\)-parton one-loop matrix element. This contribution is free from kinematic singularities (the infrared-safe jet function screens any remaining divergences of the squared matrix elements), but it contains explicit \(\epsilon\)-poles which come from integrations over loop momenta. The idea behind the method is to define completely local subtraction terms for the NNLO correction in the same spirit as was done at NLO accuracy. Thus the \(m+2\)-parton contribution is made finite by introducing local subtraction terms whose kinematic singularities exactly reproduce those of the double real emission matrix elements (in \(d\) dimensions) in each single and double unresolved limit: \[\begin{aligned}
\sigma^{\rm NNLO}_{m+2}[J] = \int_{m+2}\bigg\{& {\rm d}\sigma_{m+2}^{\rm RR}J_{m+2}-{\rm d}\sigma_{m+2}^{\rm RR,A_2}J_{m} \notag\\&-\bigg[{\rm d}\sigma_{m+2}^{\rm RR,A_1}J_{m+1}-{\rm d}\sigma_{m+2}^{\rm RR,A_{12}}J_{m} \bigg] \bigg\}_{d=4} \. \label{eq:sigNNLOm+2} \end{aligned}\] In , \({\rm d}\sigma^{\rm RR,A_2}_{m+2}\) regularizes those singularities of the RR contribution which emerge in double unresolved limits, while \({\rm d}\sigma^{\rm RR,A_1}_{m+2}\) serves as a local counterterm for single unresolved singularities. The last term, \({\rm d}\sigma^{\rm RR,A_{12}}_{m+2}\), is introduced to remove both the kinematic singularities that develop in \({\rm d}\sigma^{\rm RR,A_2}_{m+2}\) in single unresolved regions and also the singularities of \({\rm d}\sigma^{\rm RR,A_1}_{m+2}\) in double unresolved ones. The precise definitions of all subtraction terms that appear in were given in. The \(m+1\)-parton contribution takes the form \[\begin{aligned}
\sigma^{\rm NNLO}_{m+1} = \int_{m+1}\bigg\{& \bigg( {\rm d}\sigma_{m+1}^{\rm RV} + \int_1{\rm d}\sigma^{\rm RR,A_1}_{m+2} \bigg)J_{m+1} \notag \\&-\bigg[ {\rm d}\sigma_{m+1}^{\rm RV,A_1} + \left(\int_1{\rm d}\sigma_{m+2}^{\rm RR,A_1}\right)\!\strut^{\rm A_1} \bigg]J_{m} \bigg\}_{d=4} \,, \label{eq:sigNNLOm+1} \end{aligned}\] where the first line of contains the RV contribution as well as the integrated form of the single unresolved subtraction term in , \(\int_1{\rm d}\sigma^{\rm RR,A_1}_{m+2}\). The sum of these two terms is free of \(\epsilon\)-poles, however both terms still contain kinematical singularities when a parton becomes unresolved. These singularities are regularized by the local subtraction terms on the second line of . The exact definitions of these subtraction terms were presented in. The last contribution to the NNLO correction is the \(m\)-parton one which contains the VV contribution along with the integrated forms of all remaining subtraction terms which we have not yet added back. Schematically this can be written as \[\begin{aligned}
\sigma^{\rm NNLO}_{m} = \int_{m}\bigg\{& {\rm d}\sigma_{m}^{\rm VV} + \int_{2}\bigg[{\rm d}\sigma^{\rm RR,A_2}_{m+2}-\sigma^{\rm RR,A_{12}}_{m+2}\bigg] \notag\\& + \int_1\bigg[ {\rm d}\sigma^{\rm RV,A_1}_{m+1} + \bigg(\int_1{\rm d}\sigma^{\rm RR,A_1}_{m+2}\bigg)\!\strut^{\rm A_1} \bigg] \bigg\}_{d=4}J_{m} \. \label{eq:sigNNLOm} \end{aligned}\] Since the \(m+2\)-and \(m+1\)-parton contributions in are both finite by construction, the finiteness of the \(m\)-parton piece in is automatic and guaranteed by the KLN theorem. The various integrated approximate cross sections that appear in above were computed in a series of papers, culminating in the explicit demonstration of the finiteness of this contribution for electron-positron annihilation into three jets in . Since the subtractions render all three contributions finite, they can be separately integrated numerically using standard Monte Carlo techniques. We stress that since all subtractions are completely local, the integrations may be performed with any convenient numerical procedure.
# Electron-positron annihilation into three jets
We have implemented the scheme as outlined above into the `fortran90` program library (Monte Carlo for the Subtraction Method). The implementation is completely general for processes with colorless initial states, with only the squared matrix elements for a given process (including the color-and spin-correlated ones) as necessary inputs. As a first application, we used our method and code to compute NNLO QCD corrections to physical observables in three-jet production in electron-positron annihilation . Since these corrections are known for several quantities in the literature , (see also which describes the program implementing the computations of ) this provides an excellent opportunity to validate our method and the framework implementing it. Hence, we compared our predictions for the six standard event shape variables of thrust (\(T\)), C-parameter, total-and wide-jet broadening, heavy-jet mass and the two-to-three jet transition variable \(y_{23}\) in the Durham jet clustering algorithm to the predictions of [@GehrmannDeRidder:2007hr; @Weinzierl:2009ms]. We performed the comparisons at the LEP2 energy of \(\sqrt{s} = m_Z = 91.2\gev\). The perturbative coefficients were defined using the normalization common at lepton-lepton colliders: \[\begin{aligned}
\frac{O}{\sigma_0}\frac{{\rm d}\sigma}{{\rm d}O} = \frac{\alpha_{\rm s}}{2\pi}O\,A(O) + \left(\frac{\alpha_{\rm s}}{2\pi}\right)^2 O\,B(O) + \left(\frac{\alpha_{\rm s}}{2\pi}\right)^3 O\,C(O) + \mathcal{O}(\alpha_{\rm s}^4) \,, \end{aligned}\] where \(\sigma_0\) is the leading order cross section for \(e^+\,e^-\to\,{\rm hadrons}\) and \(O\) is the event shape variable for which we obtain the NNLO accurate prediction. We present the comparison of our predictions to those of (denoted by GGGH[^1]) and Ref. (denoted SW[^2]) for two representative cases, thrust (\(\tau = 1-T\)) and the C-parameter in and . In these plots, the left hand panels show the physical predictions for the observables at LO, NLO and NNLO accuracy, together with the data measured by the ALEPH collaboration. The bands correspond to scale variations in the range of \([m_Z/2,2m_Z]\) with \(\mu_0 = m_Z\) chosen as the central scale. While these plots clearly show the convergence of the perturbative series for both the \(\tau\) and C-parameter distributions as we go from LO to NLO and NNLO accuracy, the comparison with data also makes it evident that parton shower and non-perturbative corrections are sizable. On the lower panels we plot the ratios of the predictions of GGGH (bottom panel) and SW (middle panel) normalized to our results and find agreement between the various computations, except at the kinematic limits of the distributions. In order to better quantify the level of agreement among the perturbative predictions, on the right hand panels of and we present the comparisons of the NNLO coefficients directly. We plot the distribution of the NNLO coefficient \(O\, C(O)\) on the top panels, while the middle and bottom panels again show the ratios of the predictions of SW and GGGH normalized to our results. The narrow gray bands on the middle and lower panels show the numerical uncertainty of our computation due to Monte Carlo integrations. We observe a very good numerical convergence of our method at NNLO. Examining these plots, we see that the agreement is generally quite good between the predictions of SW and and reasonably good between GGGH and , with the precise comparison to GGGH being hampered by the somewhat large numerical uncertainties of those predictions. We also see that significant deviations are present for small and large values of the event shapes. For example the differences between and the two other predictions grow up to a factor of two for \(\tau > 1/3\). However, in this region the contribution from the three-particle final state vanishes and the thrust distribution is determined by a four-jet final state. Hence, in this region \(C(\tau)\) is given by NLO corrections to four-jet production, which have been known for a long time and can also be computed by modern automated tools such as Ẇe find that our predictions are in complete agreement with those for the thrust distribution for \(\tau>1/3\), and with the computation of Ref. for the C-parameter distribution for \(C_{\mathrm{par}}>3/4\). For small values of the event shapes, we checked that our predictions are in agreement with the resummed computations obtained from SCET expanded to \(\mathcal{O}(\alpha_{\rm s}^3)\). Beside the standard event shape variables discussed above, we computed for the first time predictions at NNLO accuracy for oblateness, energy-energy correlation (EEC) and jet cone energy fraction (JCEF). Here we present our results for jet cone energy fraction, which is defined as \[\begin{aligned}
\frac{{\rm d}\Sigma_{\rm JCEF}}{{\rm d}\cos\chi} &= \sum_{i}\int{\rm d}\sigma_{e^+\,e^-\to i+X} \frac{E_i}{Q}\delta(\cos\chi + \cos\theta_{ij}) \delta\left( \cos\chi-\frac{\vec{p}_i\cdot\vec{n}_T}{|\vec{p}_i|} \right) \,, \end{aligned}\] where \(Q\) is the center of mass energy, \(E_i\) is the energy of particle \(i\) (in the center of mass frame), \(\cos\theta_{ij}\) is the cosine of the angle between the three-momenta of particles \(i\) and \(j\) (also in the center of mass frame) and \(\vec{n}_T\) is the thrust axis pointing from the heavy to the light jet mass hemisphere. Our physical predictions for the jet cone energy fraction at LO, NLO and NNLO accuracy together with our prediction for the NNLO coefficient \(C_{\mathrm{JCEF}}(\chi)\) are presented in . Our code displays a good numerical convergence also for these distributions. On the left hand panel showing the physical prediction, we have also included experimental data measured by the DELPHI collaboration. We observe that perturbative corrections are rather small over a wide range of angles. Hadronization corrections and detector corrections for this observable are also known to be quite small and indeed the perturbative result alone is seen to give a rather reasonable description of the data. Hence, jet cone energy fraction is a particularly simple and promising observable for the precise extraction of from data. Finally, we turn to the computation of jet rates at NNLO accuracy. The production rate for \(n\)-jet events in electron-positron annihilation is given by the ratio of the \(n\)-jet cross section to the total hadronic cross section: \[\begin{aligned}
R_{n}(y_{\mathrm{cut}}) = \frac{\sigma_{n}(y_{\mathrm{cut}})}{\sigma_{\mathrm{tot}}}\. \end{aligned}\] Here the \(n\)-jet cross section \(\sigma_{n}(y_{\mathrm{cut}})\) must be defined using an infrared-safe jet clustering algorithm. One class of such algorithms are the exclusive sequential recombination algorithms. Here we focus on the Durham algorithm for which the resolution variable is defined as \[\begin{aligned}
y_{ij} = \frac{2\min(E_i^2,E_j^2)(1-\cos\theta_{ij})}{Q^2} \end{aligned}\] and recombination is performed in the E-scheme, i.e., the four-momenta of the objects to be combined are simply added. We present our preliminary physical predictions for \(R_{3}(y_{\mathrm{cut}})\) in the Durham clustering algorithm in the left panel of at LO, NLO and NNLO accuracy, together with measured data from the OPAL collaboration. Although the inclusion of the NNLO corrections vastly improves the theoretical description of the measurement, a sizable difference remains between the NNLO prediction and the data for small values of the resolution parameter. We attribute this difference to missing parton shower (or resummation) and hadronization effects. In order to improve the situation, we match our perturbative prediction to the resummed result, which is known for \(R_{3}(y_{\mathrm{cut}})\) in the two-jet limit (i.e., \(y_{\mathrm{cut}} \to 0\)) up to next-to-leading-logarithmic (NLL) accuracy. (The distribution of the two-to-three jet transition variable \(y_{23}\) in the Durham algorithm is known up to NNLL accuracy.) This matching is performed as follows. We write the fixed order prediction for the three-jet rate at NNLO accuracy as \[\begin{aligned}
R_3^{\mathrm{NNLO}}(y_{\mathrm{cut}}) = \frac{\as}{2\pi} A_3(y_{\mathrm{cut}}) + \left(\frac{\as}{2\pi}\right)^2 B_3(y_{\mathrm{cut}}) + \left(\frac{\as}{2\pi}\right)^3 C_3(y_{\mathrm{cut}})\. \label{eq:R3NNLO} \end{aligned}\] The resummed prediction at NLL accuracy can be written in the following form, \[\begin{aligned}
R_3^{\mathrm{NLL}}(y_{\mathrm{cut}}) = 2 [\Delta_q(Q)]^2 \int_{Q_0}^Q \mathrm{d} q\, \Gamma_q(Q,q) \Delta_g(q)\,, \label{eq:R3NLL} \end{aligned}\] where \(Q_0 = \sqrt{y_{\mathrm{cut}}} Q\) and the \(\Delta_i(Q'')\) and \(\Gamma_i(Q'',Q')\) functions are given explicitly in up to NLL accuracy. When evaluating numerically, we use the one-loop running of in \(\Delta_i(Q'')\) and \(\Gamma_i(Q'',Q')\). In order to perform the matching, we expand in powers of up to and including \(\mathcal{O}(\alpha_{\mathrm s}^3)\) terms: \[\begin{aligned}
R_3^{\mathrm{NLL}}(y_{\mathrm{cut}}) &= \frac{\as}{2\pi} A_3^{\mathrm{NLL}}(y_{\mathrm{cut}}) + \left(\frac{\as}{2\pi}\right)^2 B_3^{\mathrm{NLL}}(y_{\mathrm{cut}}) + \left(\frac{\as}{2\pi}\right)^3 C_3^{\mathrm{NLL}}(y_{\mathrm{cut}}) \notag\\& + \mathcal{O}(\alpha_{\rm s}^4)\. \label{eq:R3NLLexp} \end{aligned}\] Our final expression at NNLO+NLL accuracy is then given by \[\begin{aligned}
R_3^{\mathrm{NNLO+NLL}}(y_{\mathrm{cut}}) = R_3^{\mathrm{NLL}}(y_{\mathrm{cut}}) &+ \frac{\as}{2\pi} \Big[A_3(y_{\mathrm{cut}})-A_3^{\mathrm{NLL}}(y_{\mathrm{cut}})\Big] \notag\\& + \left(\frac{\as}{2\pi}\right)^2 \Big[B_3(y_{\mathrm{cut}})-B_3^{\mathrm{NLL}}(y_{\mathrm{cut}})\Big] \notag\\& + \left(\frac{\as}{2\pi}\right)^3 \Big[C_3(y_{\mathrm{cut}})-C_3^{\mathrm{NLL}}(y_{\mathrm{cut}})\Big]\. \label{eq:R3NNLO+NLL} \end{aligned}\] The right hand panel of shows the preliminary results of this matching procedure. We indeed see a marked improvement of the theoretical description, together with a significant reduction in the relative scale uncertainty below \(y_{\mathrm{cut}} \sim 10^{-2}\). Since jet rates computed using different jet algorithms can have different sensitivities to non-perturbative effects, it would be interesting to extend these results to other jet clustering algorithms as well.
# Conclusions
In this contribution we briefly outlined the subtraction method for computing NNLO QCD corrections for processes with colorless initial states. As a first application, the method was used to compute physical observables in three-jet production in electron-positron annihilation. After validating our numerical program by comparisons to existing computations, we presented NNLO QCD results for jet cone energy fraction, which has not been computed at NNLO accuracy before. We find that the perturbative corrections for this observable are rather small for a wide range of angles. This, together with the smallness of the hadronization and detector corrections makes the jet cone energy fraction a very promising observable for the precise extraction of the strong coupling from data. Finally, we presented preliminary results for the three-jet rate in the Durham algorithm at NNLO and NNLO+NLL accuracy.
[^1]: We are grateful to G. Heinrich for providing the predictions of for us.
[^2]: In these comparisons we use updated (with respect to those published in Ref. ) but unpublished predictions provided to us by S. Weinzierl. We are grateful to S. Weinzierl for providing these updated results for us. | {'timestamp': '2017-06-07T02:05:06', 'yymm': '1706', 'arxiv_id': '1706.01688', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01688'} |
# Introduction
Blue compact dwarf galaxies (BCDs) are low-mass systems characterized by active star formation in compact regions and strong narrow emission lines superposed on a flat stellar continuum. They were originally considered as either nascent or immature galaxies that had recently undergone intense bursts of star formation. However, it turns out that they also contain stellar populations older than \(\sim\)`<!-- -->`{=html}1 Gyr as shown by a number of optical and near-IR studies, implying that BCDs are little galaxy fossils where starbursts have been recently triggered. The low metallicity of BCDs also suggests that they have not always maintained their current star formation rate. Indeed, the neutral hydrogen gas mass found in very gas-rich BCDs is typically \(10^8M_\odot\), which will require 1 Gyr to be consumed entirely if they keep up with the current star formation rate. Instead, a number of models have shown that dwarf galaxies may go through a "BCD phase" due to the gas infall, merging with another dwarf galaxy, or fly-by interaction. Those scenarios have been also tested observationally in a wide range of wavelengths from optical through to radio. For example, merging signatures or the presence of companions have been confirmed in H\(\alpha\) and deep optical imaging studies. In addition, Hi imaging has been particularly useful in finding the evidence for tidal interactions in BCDs. Approximately 57% of BCDs have been found with gas-rich but optically faint dwarf nearby companions, and the fraction increases up to 80% when BCDs with signatures of tidal interaction including merging are included. All these observations support the hypothesis that the bursts of star formation in BCDs can be induced externally. On the other hand, there are cases of BCD pairs with separation larger than a few tens kpc such as UM 461/462, ESO 338-004/004B, ESO 400-043/043B, and IRAS 08339+6517/2MASX J08380769+6508579. For those BCD pairs, it is more difficult to prove that the two galaxies are or have been tidally interacting with each other. And even if this is so, how their \"fly-by\" interaction could lead to the recent active star formation is still questionable. Considering that there is quite a high fraction of BCD galaxies that are not in the process of merging, it is important to show that they can indeed originate from fly-by interactions. A pair of BCDs, ESO 435\(-\)IG 020 and ESO 435\(-\) G 016, separated by \(\sim\)`<!-- -->`{=html}80 kpc in projection at similar distances, is one such example. Based on the fact that both galaxies appear to be highly disturbed in the optical observations, have classified this pair as "detached interacting BCDs", proposing that their peculiar optical morphology originates from tidal interaction with each other. Intriguingly, a large neutral hydrogen gas envelope has been revealed in the Hi Parkes all sky survey, which is found to be covering both galaxies (see Figure [\[fig_hipass\]](#fig_hipass){reference-type="ref" reference="fig_hipass"}). Their optical morphologies and the common Hi envelope make this pair a good candidate for the case of a BCD pair with fly-by interaction origin. Considering the distance between the two galaxies, however, it is difficult to conclude whether Hi gas associated with individual galaxies simply appears to be connected due to the large HIPASS beam, which is comparable to the separation of the pair on the sky, or whether they are indeed an interacting pair. In order to test the hypothesis that 1) these two dwarfs are going (or have gone) through a fly-by interaction, and 2) their interaction is the cause of their recent star formation activity, we followed up this region in Hi with a higher resolution using the Australia Telescope Compact Array (ATCA). In this work, we probe detailed Hi morphology and kinematics of the ESO 435\(-\)IG 020/016 pair and their star formation properties to investigate the role of a fly-by interaction as one of the important origins for quite a large fraction of the BCD population including our targets. This paper is organized as follows. In Section 2, general properties of ESO 435\(-\) G 016/020 pair are introduced. In Section 3, the ATCA Hi observation and data reduction are described. In Section 4, the Hi morphology and kinematics of individual galaxies are presented. In Section 5, the presence of intergalactic Hi between the pair is inspected. Also, the star formation and interaction histories of the pair are discussed. In Section 6, we summarize and conclude.
**Notes.** Information on the galaxy coordinates (R.A., DEC.) and radial velocities comes from. \(B-\)band luminosities, stellar masses, optical extents, and distances are taken from, and. Distances are corrected for the relative motion to the centroid of the Local Group.
**Notes.** The 2nd and 3rd columns (\(\Delta_{\rm RA}\) and \(\Delta_{\rm Dec}\)) indicate the offset between the optical center and the peak of the Hi emission. For Hi mass, the distances in Table [\[tbl_genpro\]](#tbl_genpro){reference-type="ref" reference="tbl_genpro"} have been applied. The Hi velocity defined within the symmetric part of the gas disk is given within the parenthesis.
[^1]: Astronomical Image Processing System, *http://www.aips.nrao.edu*
[^2]: Groningen Image Processing System,\
`https://www.astro.rug.nl/\tilde{~}gipsy/` | {'timestamp': '2017-06-08T02:00:39', 'yymm': '1706', 'arxiv_id': '1706.01902', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01902'} |
# Introduction
The energy transfer in the light-harvesting antenna complex (LHC) takes place via exciton propagation among pigments bound to the LHC proteins. The exciton is created by resonant photo-absorption on an antenna pigment, leading to electron excitation from the ground to the excited energy level, \(\gamma +E_0 \to E_1\), Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. Due to the dipole-dipole interaction, \(V\), the exciton then propagates between neighboring pigments to the reaction center (RC), while all, excited and non-excited sites of the antenna, remain uncharged. Finally, the exciton arrives at the site \(N\) (the "donor" of the RC), where the primary process of the charge (electron) separation occurs. As a result the, donor becomes positively charged, and the electron participates in chemical reactions in the RC. Finally, at some time \(\tau\) the donor's charge is reduced by an electron ultimately derived from water oxidation, Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}, and the cycle is completed. The dynamics of the exciton transfer along the antenna, including the primary charge separation is very rapid (\(\sim\) ps). Otherwise, the exciton would be lost by fluorescence or by other (recombination) processes, taking place on the time-scale of \(\sim\) ns. In comparison, the duration of entire cycle (\(\tau\)), completed with reduction of the primary oxidized donor, is much longer (\(\sim \mu s\)). During the cycle no excitons occupy the RC donor. However, they can be accumulated by the antenna pigments, and finally being lost by fluorescence, as shown schematically in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. Usually, the study of energy-transfer in the LHC is limited by a single-exciton migration along the antenna pigment bed. The initial photo-absorption, fluorescence and charge restoration of the RC donor are not included in the consideration, since they occur at significantly different time-scales. Nevertheless, without a consistent accounting of all these multi-scale processes one cannot fully understand and describe the exciton dynamics in the LHC. For instance, for efficient performance of the LHC, no more than one photon can be absorbed during the RC cycle, \(\tau\), corresponding to the optimal absorption rate, \(\sim 1/\tau\). However, if the intensity of the sunlight increases, more than one exciton can accumulate inside the antenna. These "trapped\" excitons can damage the photosynthetic apparatus through de-excitation pathways leading to the generation of oxygen singlets. In competition with this destructive pathway is de-excitation of excess excited states by non-photochemical quenching (NPQ) mechanisms. It is clear that, in order to understand the NPQ mechanism, energy transfer from the LHC to the RC cannot be treated in terms of a single-exciton transport, during one cycle. At first sight, exciton transport along the LHC appears similar to spinless electron transport in a mesoscopic system. Indeed, no more than one exciton can reside on the same site, if only one excitation is allowed for each site (hard exciton model). As a result, the exciton propagation along the antenna would be similar to electron tunneling through coupled-dot system, as shown schematically in Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}. Here, the left and the right leads (source and drain), with tunneling rates, \(\Gamma_{L,R}\), play the role of the initial photon flux and the RC, Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}, Note, that one can assume that \(\Gamma_L\ll \Gamma_R\), so that no more than one electron enters the system (similar to the LHC exciton transport in normal regime). The electron would make coherent oscillation between the dots, unless it arrives to the drain (sink). Then, it cannot return back, since the number of available states in the sink is very large. Also, the probability of its return to the source lead would be negligible with respect to the incoming flux of electrons. This also could explain the irreversibility of the exciton current. Therefore, such an analogy between the exciton and electron transport can be helpful. One can also introduce a coupling of each dot with another (leakage) reservoir which would be analogous to a loss of excitons due to fluorescence in the LHC. The treatment of electron current through the coupled dots, Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}, can be greatly simplified in the case of a large bias voltage \((\mu_L-\mu_R\gg\Gamma, V)\). Then, the many-body Schrödinger equation \(i|\dot\Psi(t)\rangle =H|\Psi(t)\rangle\), where \(H\) is the total Hamiltonian, can be reduced to the Lindblad-type particle-number-resolved Master equations, beyond the commonly used weak coupling approximation, see Ref. \[\] and recent review paper \[\], with detailed derivations. Using this method, one can evaluate in a simple way the electron current through the system, together with the leakage electron current. It is desirable to realize this analogy and derive similar Master equations for the exciton transport in the LHC, Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. However, in this case we have to include in the Hamiltonian, \(H\), terms describing the primary charge separation, which takes place at the RC donor site, and the restoration of the primary donor's neutrality. The primary charge separation can be considered as an electron transition from the excited donor state, \(E_N\), to the acceptor (charge) states, \(E_{C_i}\). The process follows by the charge stabilization, which implies that the electron is not returning to the initial donor state, \(E_N\). Therefore, the acceptor cannot be represented by a single state. Otherwise, the electron would oscillate between the donor and the acceptor. For this reason we have to consider the acceptor to be a band of dense levels, \(E_{C_i}\) (sink), centered around \(E_N\). Then, the primary charge separation is represented by tunneling of the electron from the donor to the sink with a rate, \(\Gamma\sim\)`<!-- -->`{=html}1/ps. In order to describe the restoration of the primary donor's neutrality in a phenomenological way, it is not necessary to know all details of the slow chemical reactions in the RC, initiated by the electrons. What is relevant, is only the period of the cycle (\(\tau\)). Thus, we can effectively represent the charge restoration as a slow relaxation of an electron from the acceptor to the donor's ground state, accompanied by the emission of the energy (\(E_N-E_0\)) in the RC, Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. Although the cycle is completed by an electron coming from a different place (such as water splitting in the photosystem II, etc.), its origin is not important for the exciton dynamics in the LHC. The relaxation process cannot be described by considering the sink (\(E_{C_i}\)) as only an effective Markovian reservoir, absorbing an electron after the primary charge separation. Indeed, the electron cannot return from the reservoir to the (localized) ground state with large energy transfer. Rather, the relaxation can be modeled by emission of a fictitious particle (boson) carrying this energy. For this reason, we introduce in the total Hamiltonian a very weak coupling of each electron state in the sink with a bath of fictitious bosons. If the bath is initially empty, the electron would exponentially decay to the donor's ground state with a relaxation decay rate, \(\gamma_R=1/\tau\), by emitting fictitious bosons with the energy, \(E_{N}-E_0\), Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. Inclusion of a fictitious bosons together with a quantized electro-magnetic field, describing the initial light (photon) source and fluorescence, into the total Hamiltonian, would allow us to derive closed Master equations for exciton dynamics in a complete quantum mechanical way. As a result, we would be able to evaluate the energy (exciton) current as a function of the incoming sunlight intensity and the probability of the multi-exciton states inside the antenna. Dissipation of these multi-excitons by fluorescence is studied as well, by evaluation of the fluorescent current. This approach can be considered as a framework for the treatment of the energy transfer through any network in the antenna complex and also in the presence of noise. The latter, however, will be studied in a separate work. The paper is organized as follows. In Sec. IIA, we present a derivation of the Master equations for photo-absorption by a single electron site, using our wave-function approach. In this way, we reproduce old results, existing in the literature, but obtained in a different way. Primary charge separation and restoration on the donor site is investigated in Sec. IIB. Section IIC presents the Master equations, describing a complete cycle for one site antenna. Simple analytical expressions for the energy current and for the probability of different exciton states are obtained. Sec. III deals with the general case of an \(N\)-site antenna. The Hamiltonian and the general particle-number-resolved Master equations are presented in Sec. III A. The examples of two and three-site antenna are discussed in detail in Sec. III B,C. A particular attention is paid to the fluorescent current and to its dependence on the sunlight intensity. Last section is the Summary.
# Derivation of Master equations for one-site antenna
## Rate equations for photo-absorption
Let us derive Master equations for photo-absorption on the first (peripheral) site of antenna, Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}, separated from the rest of antenna. The corresponding Hamiltonian, \(H_1\), can be written as, \[\begin{aligned}
H_1=E_0\hat a_{0}^\dagger \hat a_{0}^{}&+E_1\hat a_1^\dagger \hat a_1^{}+\sum_k\omega_k\hat C_k^\dagger \hat C_k^{}\nonumber\\ &~~~~~~+\sum_k\big(g_k\hat B_1^\dagger\hat C_k+H.c.\big), \label{apb1} \end{aligned}\] where \(\hat a_1^\dagger(\hat a_1^{})\) is an electron creation (annihilation) operator for the excited state, \(E_1\), and \(\hat a_{0}^\dagger (\hat a_{0}^{})\) is the same for the ground state, \(E_0\), (in following we take \(E_0=0\)), while \(\hat C_k^\dagger (\hat C_k^{})\) is a photon creation (annihilation) operator. \(\hat B_1^{\dagger}=\hat a_1^{\dagger}\hat a_{0}^{}\) denotes an exciton creation operator. The last term in ([\[apb1\]](#apb1){reference-type="ref" reference="apb1"}) describes the electron-photon interaction in the rotating-wave approximation. The time-dependent wave-function of the system is obtained from the Scrödinger equation, \(i\partial_t|\Psi (t)\rangle =H_1|\Psi (t)\rangle\), where the initial state of the photon bath is a mixture of pure states, \[\begin{aligned}
|\bar 0\rangle =\prod_{\bk}{(\hat C_{\bk}^\dagger)^{n_{\bk}^{}}\over\sqrt{n_{\bk}^{}!}}\hat a_0^\dagger|0\rangle\equiv \prod_{\bk}\hat a_0^\dagger|n_{\bk}\rangle, \label{apb2} \end{aligned}\] and the index \(\bk\) denotes the initially occupied states of the photon bath. Note that electron relaxation from higher to lower level, generated by the rotated-wave Hamiltonian ([\[apb1\]](#apb1){reference-type="ref" reference="apb1"}), is accompanied by one-photon emission, while the reverse process is accompanied by one-photon absorption. As a result, the many-particle wave function, describing the evolution of the entire system can always be represented as, \[\begin{aligned}
&|\Psi (t)\rangle = \Big [ b_0(t) + \sum_{\bk} b_{1\bk}(t) {\hat B_{1}^{\dagger}\hat C_{\bk}^{}\over\sqrt{n_{\bk}^{}}} + \sum_{\bk,k} b_{\bk k}(t){\hat C_{k}^{\dagger}\hat C_{\bk}^{}\over\sqrt{n_{k}^{}+1}\sqrt{n_{\bk}^{}}}\nonumber\\ &+ \sum_{\bk <\bk',k} b_{1\bk\bk' k}(t) { B_{1}^{\dagger}\hat C_{k}^{\dagger}\hat C_{\bk}^{}\hat C_{\bk'}^{}\over\sqrt{n_{k}^{}+1}\sqrt{n_{\bk}^{}}\sqrt{n_{\bk'}^{}}} + \ldots \Big ] |\bar 0\rangle, \label{wf} \end{aligned}\] where, \(b_0(t)\), \(b_{1\bk}(t)\), \(b_{\bk,k}(t)\), and \(b_{1\bk\bk' k}(t)\), are the probability amplitudes to find the system in the corresponding state, defined by the associated creation and annihilation operators. Here \(\bk, k\) denote the states of absorbed and emitted photons. The initial conditions corresponds to \(b_0(0)=1\), and all the other \(b(0)\)'s being zeros. Substituting Eq. ([\[wf\]](#wf){reference-type="ref" reference="wf"}) into the time-dependent Schrödinger equation, \(i\partial_t|\Psi (t)\rangle =H_1 |\Psi (t)\rangle\), we find an infinite set of coupled equations for amplitudes, \(b(t)\),
For simplicity, we assumed that the photon-electron coupling, \(g_k\equiv g(\omega_k)\), is real. Using the Laplace transform, \[\begin{aligned}
\tilde b(E)=\int\limits_0^\infty b(t)e^{iEt}dt, \label{laplace} \end{aligned}\] one can rewrite Eqs. ([\[apb4\]](#apb4){reference-type="ref" reference="apb4"}) as,
where the r.h.s. reflects the initial condition. To simplify these equations, we use the same technique as developed in Refs. \[\] for the electron transport through quantum dots. It consists in a replacement of all sums by integrals, \(\sum_k\to\int\rho(\omega_k) d\omega_k\), where \(\rho\) is the density of photon states. Then, substituting the amplitude, \(b(E)\), in each of the sums by an expression obtained from resolving the subsequent equation, we can perform an integration over the photon energy (\(\omega_k\)) analytically in the case of a wide band-width limit (Markovian reservoir). This implies that the spectral function, \(|g(\omega_k)|^2\rho(\omega_k)\), is weakly dependent on \(\omega_k\) and \(E_1\gg g^2\rho\). However, no weak coupling limit is needed. The procedure is almost identical to that described in details in Sec 2 of Ref. \[\], providing that electrons are replaced by photons. The latter results in the additional factors (\(\sqrt{n_k^{}}\) and \(\sqrt{n_k^{}+1}\)) in front of the coupling constant, \(g_k^{}\). Proceeding with the same algebra, as in Ref. \[\], and then performing the inverse Laplace transform, \[\begin{aligned}
b_{0,1}(t)=\int\limits_{-\infty}^\infty \tilde b_{0,1}(E)e^{-iEt}{dE\over 2\pi}, \label{invlap} \end{aligned}\] we transform Eqs. ([\[apbb4\]](#apbb4){reference-type="ref" reference="apbb4"}), to the following equations,
where \[\begin{aligned}
\Gamma_{in}=\bar n\gamma~~{\rm and}~~\Gamma_{out}=(\bar n+1)\gamma, \label{gammas} \end{aligned}\] with \(\gamma=2\pi g^2(E_1)\rho (E_1)\), and \(\bar n=n(E_1)\). Thus, \(\Gamma_{in}\) is a rate of a photo-absorption, leading to electron transition from the ground to the excited state (exciton creation), and \(\Gamma_{out}\) is a rate of a photo-emission in a reverse process (exciton annihilation). Equations ([\[apbb5\]](#apbb5){reference-type="ref" reference="apbb5"}) can be transformed to equations for the exciton density matrix, \(\sigma(t)\), defined as, \[\begin{aligned}
\sigma_{00}^{}(t) &= |b_{0}(t)|^2 + \sum_{\bk,k} |b_{\bk k}(t)|^2 + \sum_{\bk<\bk'\atop k<k'} |b_{\bk\bk' kk'}(t)|^2 + \cdots\nonumber\\ &\equiv \sigma_{00}^{(0)}(t)+\sigma_{00}^{(1)}(t)+\sigma_{00}^{(2)}(t)+\cdots, \nonumber\\[5pt] \sigma_{11}^{}(t) &= \sum_{\bk} |b_{1\bk}(t)|^2 + \sum_{\bk<\bk',k} |b_{1\bk\bk' k}(t)|^2\nonumber\\ &~~~~~~~~~~~~~~~~~~ + \sum_{\bk<\bk'<\bk''\atop k<k'} |b_{1\bk\bk'\bk'' kk'}(t)|^2 + \cdots\nonumber\\ &\equiv \sigma_{11}^{(0)}(t)+\sigma_{11}^{(1)}(t)+\sigma_{11}^{(2)}(t)+\cdots, \label{sigmas} \end{aligned}\] where, \(\sigma_{00}^{(p)}(t)\) and \(\sigma_{11}^{(p)}(t)\), are probabilities of finding the electron in the ground state and in the excited state (exciton), with \(p\) photons emitted by time \(t\). Respectively, \(\sigma_{00}(t)\) and \(\sigma_{11}(t)\) are total probabilities (\(\sigma_{00}(t)+\sigma_{11}(t)=1\)). The reduction of Eqs. ([\[apbb5\]](#apbb5){reference-type="ref" reference="apbb5"}) to rate equations for the density matrix, \(\sigma_{jj}^{(p)}(t)\), defined by Eqs.([\[sigmas\]](#sigmas){reference-type="ref" reference="sigmas"}), can be done straightforwardly (see Refs.\[\]) by multiplying each of the equations ([\[apbb5\]](#apbb5){reference-type="ref" reference="apbb5"}) by a corresponding complex conjugated amplitude, and replacing sums in ([\[sigmas\]](#sigmas){reference-type="ref" reference="sigmas"}) by integrals. As a result, we arrived to the following rate equations,
which represent the particle-number-resolved Master equations of a form,
These equations are identical to those describing the electron transport from the source to the drain through a single quantum dot, with \(\Gamma_{in}\) and \(\Gamma_{out}\) corresponding to the incoming and outgoing electron rates. Summing up these equations over \(p\), one easily finds the rate equations for total probabilities,
which can be rewritten as one equation, \[\begin{aligned}
\dot{\sigma}_{00}(t)=-(2\bar n+1)\gamma\sigma_{00}(t)+(\bar n+1)\gamma, \label{apb9a} \end{aligned}\] by taking into account that, \(\sigma_{11}(t)=1-\sigma_{00}(t)\). In the steady state limit, \(\dot\sigma_{00}(t\to\infty)\to 0\), so the ground state occupation, \(\bar\sigma_{00}=\sigma_{00}(t\to\infty)\) is, \[\begin{aligned}
\bar\sigma_{00}={\bar n+1\over 2\bar n+1}. \label{st} \end{aligned}\] Therefore \(\bar\sigma_{00}=1\) for \(\bar n=0\), corresponding to fully occupied ground state. However, with increase of \(\bar n\) (sunlight intensity), the occupation of the ground state decreases. In the limit \(\bar n\to\infty\), corresponding to high sunlight intensity, \(\bar\sigma_{00}\to 1/2\). If the photon bath is in the thermal equilibrium state, then \(\bar n=1/(e^{E_1/T}-1)\). As a result, the occupation of the ground state is, \[\begin{aligned}
\bar\sigma_{00}={1\over 1+e^{-E_1/T}}, \end{aligned}\] which is a quite known result. .
## Primary charge separation and restoration by emission of fictitious bosons.
Consider the site \(N\) ("RC donor") of antenna, where the primary charge separation takes place, Figs. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}, [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. As a result, an electron, occupying the excited level, \(E_N\), of the RC donor is transferred to a neighboring site ("acceptor"), leaving the donor positively charged. This process is very fast (\(\sim\) ps) in a comparison with the time-scales of the subsequent chemical reactions in the RC (\(\sim \mu s\)), generated by the electron. The cycle is complete, when the positively charged site, \(N\), is neutralized (reduced) by an electron. In contrast to the photo-absorption of the previous section, we treat these processes phenomenologically. The first one, corresponding to the primary charge separation and stabilization, is considered as an electron tunneling from the donor to the acceptor site. However, the latter cannot be represented as a site containing isolated levels, since then the electron cannot be trapped at the acceptor: it will periodically returning to the donor. In order to achieve the charge stabilization on the acceptor's site, we represent it as a sink, with a dense levels, \(E_{C_i}\), centered around \(E_N\). Then, the electron tunneling to the acceptor is trapped there. The corresponding tunneling rate, \(\Gamma\), is very large (\(\sim\) 1/ps). The second process is modeled by a relaxation of the electron from the acceptor's band, \(E_{C_i}\), to the donor's ground state with a very low rate, \(\gamma_R\). It takes place due to emission of fictitious bosons, representing the energy transfer to the RC. In order to describe these processes quantum-mechanically, we introduce an effective Hamiltonian, \(H_N\), for the donor, \(N\), separated from the rest of antenna, Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. It can be written as, \[\begin{aligned}
&H_N=E_N^{}\hat a_N^\dagger \hat a_N^{}+\sum_iE_{C_i}^{}\hat a_{C_i}^\dagger\hat a_{C_i}^{}+\sum_p\bar\omega_p^{}\hat F_p^\dagger \hat F_p^{}\nonumber\\ &+\sum_i\Big(\tilde V_i^{}\hat a_{C_i}^\dagger \hat a_N^{} +\sum_pf_{ip}^{}\hat a_{0}^{\dagger}\hat a_{C_i}\hat F_p^{\dagger}+H.c.\Big). \label{ham1} \end{aligned}\] Here, \(\hat a_N^\dagger\) and \(\hat a_{C_i}^\dagger\), denote electron creation operators at the site, \(N\), and at a sub-level (\(i\)) of the acceptor, \(C\). Respectively, \(\hat a_{0}^{\dagger}\equiv \hat a_{0N}^{\dagger}\), is an electron creation operator at the ground state of the donor (we chose \(E_0\equiv E_{0N}=0\)). The operator, \(\hat F_p^{\dagger}\), describes a creation of fictitious bosons, bearing the energy transferred to the RC. \(\tilde V_i^{}\) is a tunneling coupling between the donor, \(E_N\), with the sub-level, \(E_{C_i}\), of the acceptor, and \(f_{ip}^{}\) is a coupling of an electron on the acceptor with fictitious bosons. The wave function, describing a whole system (electron and fictitious bosons) can be written as, \[\begin{aligned}
|\Psi (t)\rangle &=\Big[b_N^{}(t)\hat a_N^\dagger a_{0}^{}+\sum_ib_{C_i}^{}(t)\hat a_{C_i}^\dagger\hat a_{0}^{}+\sum_p b_{0p}^{}(t)\hat F_p^\dagger\Big]|\bar 0\rangle, \label{wf0} \end{aligned}\] where \(|\bar 0\rangle\equiv \hat a_{0}^{\dagger}|0\rangle\) is the initial ("vacuum") state of the system, corresponding to empty boson bath and the electron, occupying the donor's ground state. Substituting Eq. ([\[wf0\]](#wf0){reference-type="ref" reference="wf0"}) into the Schrödinger equation, \(i\partial_t|\Psi (t)\rangle =H_N |\Psi (t)\rangle\), we find the following system of coupled equations for the amplitudes \(b(t)\),
supplemented with the initial condition, \(b_N^{}(0)=1\) and \(b_{C_i}^{}(0)=b_{0p}^{}(0)=0\). Treating these equations as in the previous section, we can perform summations in Eqs. ([\[eq0\]](#eq0){reference-type="ref" reference="eq0"}a,b) analytically, by using the continuous limit, \(\sum_i\to\int\rho_C^{}dE_{C_i}\) and \(\sum_p\to\int\bar\rho d\bar\omega_p\), where \(\rho_C^{}\) and \(\bar\rho\) are the densities of the acceptor's states, \(E_{C_i}\), and of the fictitious boson bath. Equations ([\[eq0\]](#eq0){reference-type="ref" reference="eq0"}) are treated in the same way as in the previous section, by assuming that the density of states and the spectral function, \(|\tilde V|^2\bar\rho\), and \(|f|^2\rho_C\bar\rho\), are the energy independent (Markovian limit). First, we perform the Laplace transform, \(b(t)\to\tilde b(E)\), Eq. ([\[laplace\]](#laplace){reference-type="ref" reference="laplace"}). Then, resolving Eq. ([\[eq0c\]](#eq0c){reference-type="ref" reference="eq0c"}) for the amplitude \(\tilde b_{0p}(E)\) and substituting it in the previous equation, we rewrite Eqs. ([\[eq0\]](#eq0){reference-type="ref" reference="eq0"}a,b) (after the inverse Laplace transform, Eq. ([\[invlap\]](#invlap){reference-type="ref" reference="invlap"})) as,
where, \(\Gamma=2\pi|\tilde V|^2\rho_C^{}\), is the rate of the charge separation and, \(\gamma_R^{}=(2\pi)^2|f|^2\rho_C\bar\rho\Delta=1/\tau\), is the rate of the entire cycle (charge restoration). Here, \(\Delta\simeq\Gamma\), is a width of the energy distribution of a tunneling electron in the acceptor. Both rates are phenomenological parameters, which are determined experimentally (\(1/\Gamma\sim ps\) and \(1/\gamma_R\sim \mu s\)). Now we convert Eqs. ([\[eq0c\]](#eq0c){reference-type="ref" reference="eq0c"}), ([\[eq1\]](#eq1){reference-type="ref" reference="eq1"}) into the Master equations for the density matrix of the system, defined as, \[\begin{aligned}
&\sigma_{NN}^{}(t)=|b_{N}^{}(t)|^{2},~~ \sigma_{CC}^{}(t)=\sum_i|b_{C_i}^{}(t)|^{2},\nonumber\\ &\sigma_{00}^{}(t)=\sum_p|b_{0p}^{}(t)|^2. \label{eq2} \end{aligned}\] Using the same procedure as in the previous section, we find,
Indeed, the first equation is obtained by taking imaginary part of Eq. ([\[eq1a\]](#eq1a){reference-type="ref" reference="eq1a"}), multiplied by \(b_N^*(t)\). The same procedure for Eq. ([\[eq1b\]](#eq1b){reference-type="ref" reference="eq1b"}) yields, \[\begin{aligned}
&i\dot\sigma_{CC}^{}(t) =-i\gamma_R^{}\sigma_{CC}^{}(t)+2i\tilde V\,{\rm Im}\big[b_N(t)\sum_ib_{C_i}^*(t)\big], \label{eq4} \end{aligned}\] where \(\tilde V\sum_ib_{C_i}^*(t)=i{\Gamma\over2}b_N^{}(t)\), as obtained from Eqs. ([\[eq0a\]](#eq0a){reference-type="ref" reference="eq0a"}) and ([\[eq1a\]](#eq1a){reference-type="ref" reference="eq1a"}). Substituting it into Eq. ([\[eq4\]](#eq4){reference-type="ref" reference="eq4"}), we arrive to Eq. ([\[eq3b\]](#eq3b){reference-type="ref" reference="eq3b"}). Last Eq. ([\[eq3c\]](#eq3c){reference-type="ref" reference="eq3c"}) is obtained in a similar way. Solving Eqs. ([\[eq3\]](#eq3){reference-type="ref" reference="eq3"}) for the initial condition, \(\sigma_{NN}^{}(0)=1\) and \(\sigma_{CC}^{}(0)=\sigma_{00}^{}(0)=0\), we find \[\begin{aligned}
&\sigma_{NN}^{}(t)=e^{-\Gamma t},~~\sigma_{CC}^{}(t)={e^{-\gamma_R^{} t}-e^{-\Gamma t}\over \Gamma-\gamma_R^{}}\Gamma\simeq e^{-\gamma_R^{}t},\nonumber\\ &\sigma_{00}^{}(t)=1-{\Gamma e^{-\gamma_R^{}t}-\gamma_R^{}e^{-\Gamma t}\over \Gamma-\gamma_R^{}}\simeq1-e^{-\gamma_R^{}t}. \end{aligned}\] As expected, the primary charge separation occurs during very small time-interval, \(\sim\Gamma^{-1}\), whereas the charge restoration proceeds very slowly, \(\sim\gamma_R^{-1}\).
## Complete exciton dynamics for one-site antenna.
Now we can write the Master equations for an entire cycle, by combining the exciton creation with the charge separation and restoration dynamics. We start with an example of one-site antenna. (\(N=1\)), Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}. The system is described by the Hamiltonian, \({\cal H}_1\), \[\begin{aligned}
&{\cal H}_1=E_1\hat a_1^\dagger \hat a_1^{}+\sum_iE_{C_i}^{}\hat a_{C_i}^\dagger\hat a_{C_i}^{}+\sum_k\omega_k\hat C_k^\dagger \hat C_k^{}\nonumber\\ &+\sum_p\bar\omega_p^{}\hat F_p^\dagger \hat F_p^{}+\Big[\sum_k g_k\hat B_1^\dagger\hat C_k\nonumber\\ &+\sum_i\Big(\tilde V_i^{}\hat a_{C_i}^\dagger \hat a_N^{} +\sum_p f_{ip}^{}\hat a_{0}^{\dagger}\hat a_{C_i}\hat F_p^{\dagger}\Big)+H.c.\Big], \label{ap1} \end{aligned}\] which combines two Hamiltonians considered before, Eq. ([\[apb1\]](#apb1){reference-type="ref" reference="apb1"}) and Eq. ([\[ham1\]](#ham1){reference-type="ref" reference="ham1"}), for \(N=1\). Using the same procedure as in the previous sections, we convert the total wave function to particle-number-resolved Master equations for the reduced density-matrix, \(\sigma_{jj}^{(\ell)}(t)\), where, \(\ell\), denotes a number of fictitious bosons, emitted by time \(t\) and representing a number of cycles (c.f. with Eqs. ([\[sigmas\]](#sigmas){reference-type="ref" reference="sigmas"}) and ([\[phot1\]](#phot1){reference-type="ref" reference="phot1"}) for the reduced density matrix, which includes a number of emitted photons). Note that in our previous example, Eqs. ([\[eq2\]](#eq2){reference-type="ref" reference="eq2"}) and ([\[eq3\]](#eq3){reference-type="ref" reference="eq3"}), the donor has been detached from incoming photons, so only one fictitious boson can be emitted (\(\ell=1\)). The resulting Master equations have the following form,
where \(\Gamma_{in,out}\) are given by Eq. ([\[gammas\]](#gammas){reference-type="ref" reference="gammas"}). Respectively, the total probabilities, \(\sigma_{jj}^{}(t)=\sum_{\ell}\sigma_{jj}^{(\ell)}(t)\), are given by the same Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}), traced over \(\ell\). Equations ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}) have a form of classical rate equations, although they have been derived pure quantum mechanically. These equations have a clear interpretation in terms of the "loss" and "gain" processes (borrowing the terminology of the classical Boltzmann equation). The off-diagonal matrix elements (coherences), are absent in Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}), since all transitions take place between discrete and continuous (Markovian) states. However, in the case of transitions between different discrete states, off-diagonal density-matrix elements will occur in the equations of motion. This will be illustrated in following examples. As we mentioned above, the emission of fictitious bosons, introduced phenomenologically for the restoration of the donor's neutrality, represents, in fact, the energy transfer to the RC. Therefore, Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}) can be used to evaluate the corresponding energy current to the RC. Indeed, the latter is given by \(I_{en}(t)=E_1\sum_\ell \dot P_\ell(t)\), where \(P_\ell (t)=\sum_{j}\ell \sigma^{(\ell)}_{jj}(t)\) is the probability of finding \(\ell\) emitted bosons by time \(t\). Using Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}), we easily obtain that, \[\begin{aligned}
I_{en}(t)=E_1\,\gamma_R^{}\,\sigma_{22}^{}(t), \label{cur2} \end{aligned}\] where \(\sigma_{22}^{}(t)\) is the occupation probability of the acceptor. It is determined by tracing Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}) over \(\ell\),
Consider a steady-state limit (\(t\to\infty\)), where \(\dot\sigma(t)\to0\). In this case Eqs. ([\[apbb13\]](#apbb13){reference-type="ref" reference="apbb13"}) become a system of algebraic equations for \(\bar\sigma_{jj}=\sigma_{jj}(t\to\infty)\). Taking into account that \(\bar\sigma_{00}=1-\bar\sigma_{11}-\bar\sigma_{22}\), we can easily resolve these equations, thus obtaining for the steady-state probabilities, \[\begin{aligned}
&\bar\sigma_{00}={\gamma_R^{}\,(\Gamma+\Gamma_{out})\over \Gamma\Gamma_{in}+\gamma_R^{}\,(\Gamma+\Gamma_{in}+\Gamma_{out})},\nonumber\\ &\bar\sigma_{11}={\gamma_R^{}\Gamma_{in}\over \Gamma\Gamma_{in}+\gamma_R^{}(\Gamma+\Gamma_{in}+\Gamma_{out})},\nonumber\\ &\bar\sigma_{22}={\Gamma\,\Gamma_{in}\over \Gamma\Gamma_{in}+\gamma_R^{}(\Gamma+\Gamma_{in}+\Gamma_{out})}, \label{apbb14} \end{aligned}\] where, \(\Gamma_{in,out}\), are given by Eq. ([\[gammas\]](#gammas){reference-type="ref" reference="gammas"}). In the case of \(\gamma_R =0\), one finds that \(\bar\sigma_{00}=\bar\sigma_{11}=0\), since there is no charge restoration. In the limit of \(\bar n\to\infty\) (infinitely strong light intensity), one obtains that \(\bar\sigma_{00}=\bar\sigma_{11}=\gamma_R^{}/\Gamma+{\cal O}(\gamma_R^{}/\Gamma)^2\). Respectively, \(\bar\sigma_{22}=1-2\gamma_R^{}/\Gamma+{\cal O}(\gamma_R/\Gamma)^2\). This corresponds to equal occupation of ground and exited states of the donor, \(\bar\sigma_{00}=\bar\sigma_{11}=\gamma_R/\Gamma\).
# Exciton transport in \(N\)-site antenna
## Master equations in general case.
Now we extend our treatment on the \(N\)-site antenna chain, coupled with the electromagnetic field, describing photo-absorption and fluorescence, and with fictitious boson bath, describing the donor charge restoration, Figs. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} and [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. The total Hamiltonian, describing this system can be written as, \[\begin{aligned}
{\cal H}_N&=\sum_k\omega_k\hat C_k^\dagger \hat C_k^{}+\sum_{m=1}^{N}E_m\hat B_m^\dagger \hat B_m^{}+\sum_iE_{C_i}^{}\hat a_{C_i}^\dagger\hat a_{C_i}^{} \nonumber\\ &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +\sum_p\bar\omega_p\hat F_p^\dagger \hat F_p^{}+H_{int}, \label{ham} \end{aligned}\] where \(\hat B_m^\dagger =\hat a_m^\dagger \hat a_{0m}^{}\) is an exciton creation operator on the site \(m\). Without loosing generality, we assume that the ground state energy of all sites \(m=1,\ldots N\) equals zero. All notations are the same as in Eq. ([\[ap1\]](#ap1){reference-type="ref" reference="ap1"}). The interaction term (in the rotated wave approximation) can be written as \[\begin{aligned}
H_{int}&=\sum_{m=1}^N\sum_k g_k^{}\,\hat B_m^\dagger \hat C_k^{} +\sum_{m=1}^{N-1} V_m^{}\,\hat B_{m+1}^\dagger \hat B_{m}^{}\nonumber\\ &+\sum_i\Big(\tilde V_i^{}\hat a_{C_i}^\dagger \hat a_N^{} +\sum_p f_{ip}^{}\hat a_{0}^{\dagger}\hat a_{C_i}^{}\hat F_p^{\dagger}\Big)+H.c. \label{hami} \end{aligned}\] Here the electromagnetic field is coupled with all sites of antenna. However, excitons can be generated only on the first antenna site, \(m=1\), by photon absorption. All other sites, \(m=2,\ldots N\), are coupled with empty photon reservoirs. Thus, the excitons occupying these sites can only decay by the fluorescence. Note that the exciton commutation relation , \[\begin{aligned}
=\delta_{mn}^{}(1-2\hat B_m^\dagger \hat B_n^{}), \end{aligned}\] guaranties that two or more excitons cannot occupy the same site. Therefore, the exciton motion along the antenna, describing by the Hamiltonian ([\[ham\]](#ham){reference-type="ref" reference="ham"}), ([\[hami\]](#hami){reference-type="ref" reference="hami"}), is similar to that of the spinless electron transport trough the coupled quantum dots. The corresponding Master equation can be derived from the time-dependent multi-particle Schrödinger equation, in the same way as discussed in previous examples. It represents a natural extension of Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}) and ([\[apbb13\]](#apbb13){reference-type="ref" reference="apbb13"}), and can be written as (see Eq. (75) of Ref. \[\]): \[\begin{aligned}
&\dot\sigma_{\alpha\alpha'}^{(\nu,\ell)} = i\big({\cal E}_{\alpha'}-{\cal E}_{\alpha}\big) \sigma_{\alpha\alpha'}^{(\nu,\ell)}+i\sum_{\beta}\Big (\sigma_{\alpha\beta}^{(\nu,\ell)} V_{\beta\to\alpha'}-V_{\beta\to\alpha}\sigma_{\beta\alpha'}^{(\nu,\ell)}\Big)\nonumber\\ &-{1\over2}\sigma_{\alpha\alpha'}^{(\nu,\ell)}\sum_\beta (\Gamma_{\alpha\to\beta}+\Gamma_{\alpha'\to\beta}) +\sum_{\beta,\beta'}\sigma_{\beta\beta'}^{(\nu',\ell')} \Gamma_{\beta\to\alpha,\beta'\to\alpha'}, \label{d1} \end{aligned}\] where \(|\alpha\rangle\), \(|\beta\rangle\) enumerate all *discrete* multi-exciton states in the occupation number representation, and \({\cal E}_\alpha=\sum_{m\in\alpha}E_m\) is a total energy of the state, \(|\alpha\rangle\). The upper indices, \(\nu\) and \(\ell\), in the density matrix, \(\sigma_{\alpha\alpha'}^{(\nu,\ell)}(t)\), denote the numbers of fluorescent photons and fictitious bosons emitted at time, \(t\). Note, that in the last ("gain") term, \((\nu',\ell')=(\nu-1,\ell)\) or \((\nu',\ell')=(\nu,\ell-1)\), whenever emission of fluorescence photons or fictitious bosons takes place (c.f. with Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"})). The second term in Eq. ([\[d1\]](#d1){reference-type="ref" reference="d1"}) describes the direct exciton transitions between neighboring sites, \(V_{\beta\alpha}= V_{m,m+1}\equiv V_m\), generated by dipole-dipole interaction (Förster mechanism). Obviously, this term does not appear in the rate equation ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}) for one-site antenna. One can realize that the first and second terms of Eq. ([\[d1\]](#d1){reference-type="ref" reference="d1"}) represent the commutator of the density-matrix with the Hamiltonian in the Lindbladt equation . The remaining two terms represent loss and gain processes generated by: (a) coupling of the site (\(1\)) to photon bath with rates \(\Gamma_{\alpha,\beta}\equiv\Gamma_{in,out}\), Eqs. ([\[gammas\]](#gammas){reference-type="ref" reference="gammas"}); (b) fluorescence of sites, \(m=2,\ldots,N\), with the rate \(\Gamma_{\alpha,\beta}\equiv\gamma\); (c) charge separation with subsequent emission of fictitious bosons, leading to restoration of the donor's neutrality with the rates \(\Gamma_{\alpha,\beta}\equiv\Gamma,\gamma_R^{}\), respectively, Eqs. ([\[apbb12\]](#apbb12){reference-type="ref" reference="apbb12"}). Note, that the gain processes, described by the fourth term in Eq. ([\[d1\]](#d1){reference-type="ref" reference="d1"}), are generated by the *same* exciton transition leading to decay of the states \(\beta,\beta'\) to the states \(\alpha,\alpha'\). By solving Eqs. ([\[d1\]](#d1){reference-type="ref" reference="d1"}), we can determine probabilities of any multi-exciton occupations, as well as the fluorescent current (in energy units), \(I_{fl}(t)\), and the current of energy, \(I_{en}(t)\), transferred to the RC. Those are given by (c.f. with Ref. \[\]), \[\begin{aligned}
&I_{fl}(t)=\sum_{\nu,\ell} \sum_\alpha E_{m'}\nu\dot\sigma_{\alpha\alpha}^{(\nu,\ell)}(t) =\gamma\sum_{\alpha'}E_{m'}\sigma_{\alpha'\alpha'}^{}(t), \label{cur3}\\ &I_{en}(t)=E_N\sum_{\nu,\ell} \sum_\alpha\ell\dot\sigma_{\alpha\alpha}^{(\nu,\ell)}(t) =E_N^{}\gamma_R^{}\sum_{\alpha'}\sigma_{\alpha''\alpha''}^{}(t), \label{cur3p} \end{aligned}\] where \(\sigma_{\alpha\alpha}(t)=\sum_{\nu,\ell}\sigma_{\alpha\alpha}^{(\nu,\ell)}(t)\), are total probabilities, obtained from Eqs. ([\[d1\]](#d1){reference-type="ref" reference="d1"}), and \(m'=(2,\ldots ,N)\in \alpha\) denote the fluorescent sites (the first antenna site is excluded from our definition of fluorescent current). The indices, \(\alpha'\) and \(\alpha''\), enumerate all multi-exciton states decaying by fluorescence or by emission of fictitious bosons, respectively.
## Two-site antenna.
As an example for application of Eq. ([\[d1\]](#d1){reference-type="ref" reference="d1"}), we consider exciton transport through the two-site antenna, (\(N=2\)). First, we need to enumerate all possible exciton states of the system, \(\{\alpha,\beta\}=\{0,1\ldots, 5\}\). These are shown in Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}. Note, that the exciton propagates coherently between the sites (1) and (2) due to the dipole-dipole interaction, \(V_1\). All other transitions are incoherent, where related transition rates are shown for each of the states. Now we can rewrite explicitly the Master equations ([\[d1\]](#d1){reference-type="ref" reference="d1"}) for the reduced density-matrix,
The total probabilities, \(\sigma_{\alpha\alpha'}(t)=\sum_{\nu,\ell}\sigma_{\alpha\alpha'}^{(\nu,\ell)}(t)\), are given by the same Eqs. ([\[apb14\]](#apb14){reference-type="ref" reference="apb14"}) by removing the upper indices. One can easily verify that these equations display the probability conservation, \[\begin{aligned}
\sum_{\alpha=0}^5\sigma_{\alpha\alpha}(t)=1. \label{repl00} \end{aligned}\] Solving Eqs. ([\[apb14\]](#apb14){reference-type="ref" reference="apb14"}), we can find the probabilities for one and two excitons inside the antenna, \[\begin{aligned}
&P_1(t)=\sigma_{11}(t)+\sigma_{22}(t)+\sigma_{55}(t),\nonumber\\ &P_2(t)=\sigma_{44}(t). \label{ocprob} \end{aligned}\] Using Eqs. ([\[cur3\]](#cur3){reference-type="ref" reference="cur3"}) and ([\[cur3p\]](#cur3p){reference-type="ref" reference="cur3p"}), we find that the fluorescent current (on the second site, Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}) is given by, \[\begin{aligned}
I_{fl}(t) =\gamma E_2[\sigma_{22}(t)+\sigma_{44}(t)], \label{cur44} \end{aligned}\] and the energy current to the RC is, \[\begin{aligned}
I_{en}(t) =\gamma_R^{} E_2[\sigma_{33}(t)+\sigma_{55}(t)]. \label{cur33} \end{aligned}\] Consider now the steady-state limit, \(t\to\infty\). Since in this limit \(\dot\sigma_{\alpha\alpha'}\to 0\), Eqs. ([\[apb14\]](#apb14){reference-type="ref" reference="apb14"}) become a system of algebraic equations for \(\bar\sigma\equiv \sigma (t\to\infty )\). Using the probability conservation, Eq. ([\[repl00\]](#repl00){reference-type="ref" reference="repl00"}), we can rewrite these equations as,
Steady-state probability for single exciton occupation, \(\bar P_1(\bar n)\), and steady-state currents, \(\bar I_{fl,en}(\bar n)\), (Eqs. ([\[ocprob\]](#ocprob){reference-type="ref" reference="ocprob"}), ([\[cur44\]](#cur44){reference-type="ref" reference="cur44"}) and ([\[cur33\]](#cur33){reference-type="ref" reference="cur33"}) for \(t\to\infty\)), are shown in Figs. [\[fig5p\]](#fig5p){reference-type="ref" reference="fig5p"} and [\[fig5pp\]](#fig5pp){reference-type="ref" reference="fig5pp"} as functions of \(\bar n\) (light intensity). Note, that \(\bar n\gamma=\Gamma_{in}\), Eq. ([\[gammas\]](#gammas){reference-type="ref" reference="gammas"}), is a number of photons absorbed by the first site per unit time. Here and in what follows, we choose for illustrative examples some generic values of parameters, not necessary related to a specific system, namely, \(\gamma\)=1/ns (fluorescent rate), \(\gamma_R=1/\tau=10^{-3}\gamma=1/\mu\)s (charge restoration rate), \(V_1=10^3\gamma\)=1/ ps (dipole-dipole coupling between sites), and \(\Gamma=10^3\gamma\)=1/ps (charge separation rate). Figure [\[fig5p\]](#fig5p){reference-type="ref" reference="fig5p"}, displays the probability, \(\bar P_1(\bar n)\), of trapping one exciton for two values of \(\Gamma=10^3\gamma\)=1/ps (solid line) and \(\Gamma=0\) (dashed line). The latter corresponds to no-coupling of antenna with the RC. One finds from this figure that the single-exciton occupation is saturated with increase of \(\bar n\). As expected, the saturation takes place at equal occupation of ground and excited states for \(\bar n\to\infty\), Eq. ([\[st\]](#st){reference-type="ref" reference="st"}). The probability of double-exciton occupation, \(P_2(\bar n)\), Eq. ([\[ocprob\]](#ocprob){reference-type="ref" reference="ocprob"}) is not shown here, since it is very small for \(N=2\). However, already for \(N=3\), this quantity, \(P_2(\bar n)\), becomes quite large, as demonstrated below. The steady-state energy current to the RC, \(\bar I_{en}\) (solid line, black) and the fluorescent current of the second site, \(\bar I_{fl}\) (dashed line, blue) in units of donor energy (\(E_2\)) per \(1/\gamma\)=1 ps, are shown in Fig. [\[fig5pp\]](#fig5pp){reference-type="ref" reference="fig5pp"}, as a function of \(\bar n\). One finds from this figure that the currents reach saturation already for a very small \(\bar n\), where less than one photon is absorbed at the first site during 1 \(\mu\)s. This rate is of the order of optimal photo-absorption rate, when no more than one exciton can be accumulated inside the antenna during one cycle. (The time of cycle is \(\tau=1/\gamma_R=\mu\)s.) The origin of the saturation can be easily understood. Indeed, an increase of the light intensity results in generation of additional excitons, trapped inside the antenna. These excitons will be eventually removed by fluorescence, so the energy current will not increase anymore. The time-dependent energy current, \(I_{en}(t)\), Eq. ([\[cur33\]](#cur33){reference-type="ref" reference="cur33"}), is displayed in Fig. [\[fig6pp\]](#fig6pp){reference-type="ref" reference="fig6pp"} for \(\bar n=0.1\) (solid line, black) and \(\bar n=0.01\) (dashed line, blue). One finds that time for approaching the asymptotic limit increases when the intensity of light, \(\bar n\), decreases. For an optimal intensity, fitted to a cycle period, the saturation time can be quite long. Note that the fluorescent current, \(\bar I_{fl}\), Eq. ([\[cur44\]](#cur44){reference-type="ref" reference="cur44"}), is very small, in comparison with the energy current, \(\bar I_{en}\), Fig. ([\[fig5pp\]](#fig5pp){reference-type="ref" reference="fig5pp"}). On first sight, one could expect an opposite result. Indeed, \(\bar I_{en}\) is proportional to \(\gamma_R\), whereas \(\bar I_{fl}\) is proportional to \(\gamma\) (Eqs. ([\[cur33\]](#cur33){reference-type="ref" reference="cur33"}), ([\[cur44\]](#cur44){reference-type="ref" reference="cur44"})), which greatly exceeds \(\gamma_R\). However, the fluorescent current in the case of \(N=2\) can take place only when the second site (donor) is occupied, and therefore it is very small. This, however, is not the case for \(N>2\), as will be demonstrate below.
## Three-site antenna and exciton accumulation.
Consider now a three-site antenna (\(N=3\)). All possible exciton states are shown in Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}. The state (0) displays the ground state (no excitons). The states (1),(2),(3) display one-exciton states, (5),(6),(7) display two-exciton states, and (10) displays three-exciton stares. The states (4),(8),(9),(11) are the charge states when the donor is blocked. The rate equations, describing the exciton dynamics, are obtained directly from Eq. ([\[d1\]](#d1){reference-type="ref" reference="d1"}). We find,
Here we omitted for brevity the sub-indices \(\ell\), and \(\nu\), indicating a number of fictitious bosons and fluorescence photons emitted (c.f. Eqs. ([\[apb14\]](#apb14){reference-type="ref" reference="apb14"})). Solving Eqs. ([\[apb15\]](#apb15){reference-type="ref" reference="apb15"}), we obtain the reduced density matrix of the system, \(\sigma_{\alpha\alpha'}^{}(t)=\sum_{\nu,\ell}\sigma_{\alpha\alpha'}^{(\nu,\ell)}(t)\). Then, we can evaluate probabilities for one, two and three exciton configurations inside the antenna, Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}, \[\begin{aligned}
&P_1(t)=\sigma_{11}(t)+\sigma_{22}(t)+\sigma_{33}(t)+\sigma_{88}(t) +\sigma_{99}(t),\nonumber\\ &P_2(t)=\sigma_{55}(t)+\sigma_{66}(t)+\sigma_{77}(t)+\sigma_{11,11}(t) +\sigma_{99}(t),\nonumber\\ &P_3(t)=\sigma_{10,10}(t). \end{aligned}\] The results for steady-state, \(\bar P=P(t\to\infty)\), are displayed in Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}. In contrast with the two-site antenna, Fig. [\[fig5p\]](#fig5p){reference-type="ref" reference="fig5p"}, the probability for two-exciton trapping is comparable with that for a one-exciton, \(\bar P_1(\bar n)\). Three exciton trapping, \(\bar P_3(\bar n)\), however, is negligibly small. The fluorescent current, \(I_{fl}(t)\), Eqs. ([\[cur3\]](#cur3){reference-type="ref" reference="cur3"}), from the second and the third sites, can be written explicitly as, \[\begin{aligned}
I_{fl}(t) &=\gamma[E_2\sigma_{22}(t)+E_3\sigma_{33}(t)+E_2\sigma_{55}(t)\nonumber\\ &+(E_2+E_3)\sigma_{77}(t)+E_2\sigma_{88}(t)\nonumber\\ &+(E_2+E_3)\sigma_{10,10}(t)+E_2\sigma_{11,11}(t)]. \label{cur55} \end{aligned}\] Respectively, the energy current, transferred to the RC, \(I_{en}(t)\), ([\[cur3p\]](#cur3p){reference-type="ref" reference="cur3p"}), reads, \[\begin{aligned}
I_{en}(t) =\gamma_R^{} E_3[\sigma_{44}(t)+\sigma_{88}(t)+\sigma_{99}(t)+\sigma_{11,11}(t)]. \label{cur55p} \end{aligned}\] Both currents at steady-state, \(\bar I=I(t\to\infty )\) (in units of donor energy \(E_3\) per \(1/\gamma\)=1 ns), are displayed in Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"} for aligned levels, \(E_1=E_2=E_3\), and for the same parameters as in Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}. The energy current, \(\bar I_{en}\), is displayed by solid (black) line, and the fluorescence current, \(\bar I_{fl}\), is shown by dashed (blue) line. One finds that the currents are saturated for large \(\bar n\). Here, in contrast with the case \(N=2\), Fig. [\[fig5pp\]](#fig5pp){reference-type="ref" reference="fig5pp"}, the fluorescent current strongly dominates for large \(\bar n\). However, for small \(\bar n\), the energy current dominates over the fluorescent current. Note, that in that region both currents, \(\bar I_{en,fl}(\bar n)\), are very sensitive to \(\bar n\). It implies that a reliable analysis of the optimal regime for the LHC performance cannot be done without full account of all relevant physical effects. Therefore, our microscopic approach, which provides such an account, can be very relevant to determine conditions when the LHC works with a maximal efficiency.
# Discussion
Although most investigations of energy (exciton) transport in the LHCs concentrate on one-exciton motion along the antenna, we demonstrated that it is not sufficient for a consistent description of exciton dynamics, in particular for an account of the NPQ-type processes. Therefore, we extend the Hamiltonian by including additional parts, describing the exciton creation and the fluorescence, through the interaction with the electromagnetic field, charge separation on the donor site, and the charge restoration after completing the corresponding cycle of the chemical reactions in the RC. The latter part is described phenomenologically, as an electron relaxation from the RC to the donor's ground state by emission of fictitious bosons. This process represents the energy transfer to the RC. We have to emphasize that our effective description of the charge restoration involves the relaxation of the same electron, coming from the charge separation. In reality, the process of reduction of the RC donor is more complicated. Nevertheless, this issue is not a deficiency of our model, since different electrons are indistinguishable and its origin is not relevant for the description of LHC dynamics. The relevant quantity is only the duration of the cycle, which is determined experimentally. An important assumption in our treatment, however, is the exponential relaxation process, which models reduction (charge restoration) of the RC donor. In principle, the relaxation could be of a different type, like a power-low. Then, it would imply a different Hamiltonian term describing the fictitious bosons. We consider our approach as a general framework for constructing closed Master equations for complete description of the LHC. Additional effects can be accounted for in the same way as explained above, by adding the corresponding terms in the Master equations. For instance, any site of the antenna can be coupled with a sink, modeling the NPQ processes, in addition to the fluorescence. In a similar way, one can introduce vibrational modes of the antenna sites in the Hamiltonian, together with environmental noise. These effects can play a very important role in the energy transport. Indeed, the antenna levels in general, are non-aligned. If the energy difference between sites is larger than the inter-site coupling, then transitions will be strongly reduced (Anderson localization). In this case, the quasi-resonant vibrational modes can close the gap. Perfect matching is not necessary if the environmental noise is taken into account. The inclusion of vibrational modes together with the noise in our Master equation is not considered in the present paper. This issue will be discussed in a separate work, by treating these effects in a framework of a dichotomic (telegraph) noise. Such a procedure would double the number of rate equations, without additional complications. In general, our cycled multi-exciton Mater equations represent a more detailed description of the LHC dynamics. Therefore, by using these equations one can evaluate important effects, which cannot be treated by other methods. However, the number of multi-exciton states strongly increases with a number of antenna sites, as we have already seen by comparing \(N=2\) with \(N=3\) cases. This is a general problem for any treatments involving the density-matrix. There exists, however, an alternative, single-electron approach for mesoscopic transport, which can treat the wave function, instead of the density-matrix, even in the presence of noise. Then, the number of equations will be drastically reduced. An extension of this approach to the LHC is a topic of a future investigation. | {'timestamp': '2017-06-08T02:01:56', 'yymm': '1706', 'arxiv_id': '1706.01958', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01958'} |
# Introduction {#sec:intro}
Very high energy gamma rays (VHE; E \(>\) 100 GeV) are messengers of the most energetic phenomena in the Universe. A number of pulsars, pulsar wind nebulae, supernova remnants and micro-quasars in our own Galaxy have been detected to emit such energetic radiation. Outside the Milky Way, VHE gamma rays have been seen from ultra-relativistic jets of particles escaping super-massive black holes and from galaxies with an exceptional rate of star formation. The VHE sky counts nowadays more than 150 sources. A large fraction of these sources have been detected for the fist time at VHE by the current generation of Imaging Atmospheric Cherenkov Telescopes (IACTs), namely H.E.S.S., MAGIC, and VERITAS. IACTs are designed to detect the faint Cherenkov flashes that originate from a particle shower that is formed when an energetic photon impacts the Earth atmosphere. All three telescope systems use large reflective mirrors to focus the Cherenkov (near-UV and optical) light into meter-size cameras pixelized by photomultiplier tubes. The typical signature of a gamma-ray induced shower in the camera is an image of elliptical shape and few \(ns\) duration, that has to be separated from the dominating background of mainly proton-induced showers. The H.E.S.S. array of telescopes is located in Khomas Highland in Namibia at 1800 m above sea level. The array went in operation in 2004. It is composed of four 12 m dish diameter telescopes and, since 2012, a central telescope named CT5 with a diameter of 28 m. When operated in the initial configuration of four telescopes, the system has an energy threshold of \(\sim\)`<!-- -->`{=html}100 GeV with an angular resolution better than 0.1\(^{\circ}\) and an energy resolution below 15%. The data analysis and the quality selection are described in. These performances will be boosted by a major upgrade of the cameras finished in early 2017. The performances of CT5, currently operated in mono mode, are explained in detail in. MAGIC, at the Roque de los Muchachos Observatory in the Canary island of La Palma, Spain, is a system of two telescopes operating since 2009 at 2200 m above sea level. The single telescope MAGIC-I operated in mono mode from 2004 to 2009 and was renewed in 2011/2012 to match the performances of the newer MAGIC-II. With their 17 m diameter dishes, the MAGIC telescopes reach an energy threshold as low as 50 GeV. The angular resolution, above 200 GeV is \(<\) 0.07\(^{\circ}\), while the energy resolution is 16%. The duty cycle of the system has recently been increased thanks to the possibility of taking data under partial moonlight. VERITAS is a ground-based gamma-ray instrument operating at the Fred Lawrence Whipple Observatory (FLWO) in southern Arizona, USA. VERITAS is designed to measure gamma rays with energies from \(\sim\) 85 GeV up to \(>\) 30 TeV. Since its first light observation in 2007, VERITAS underwent two major upgrades: the relocation of the first telescope in 2009 and the upgrade of the cameras in 2012. VERITAS, together with MAGIC, observes sources located in the Northern emisphere, while H.E.S.S. observes the Southern sky. Figure [\[fig:sensitivity\]](#fig:sensitivity){reference-type="ref" reference="fig:sensitivity"}, from, shows the differential sensitivity of the VERITAS and MAGIC telescopes reached with 50 hours of obervations in comparison with the sensitivity of other telescopes operating or planning to operate (e.g. LHAASO and e-ASTROGAM) in neighboring energy ranges. The figure reports also the expected sensitivity of the future Cherenkov Telescope Array (CTA), designed to bring the VHE gamma ray astrophysics in the regime of large numbers (\(>\) 1000 sources). CTA will consist of two sites, one in the Northern and one in the Southern emisphere, and tens of telescopes of three different sizes. Currently several prototypes are under construction and test. CTA is expected to start its early science in few years from now. See for a review. The duty cycle of current IACTs is mainly limited by sky brigthness and weather conditions and is below 10%. For this reason, and for the fact that the field of view is small (3.5-5\(^{\circ}\)) forcing IACTs to operate mostly in pointing mode, the H.E.S.S., MAGIC, and VERITAS collaborations have developed a dense Target of Opportunity (ToO) program. ToO observations are usually triggered by a high state of the source detected in optical and/or X-ray and/or gamma rays. Additionally, triggers by neutrino and gravitational waves are also considered. This strategy allowed the discovery of several sources during flaring states and is particularly relevant for distant emitters and for sources with a faint emission at TeV energies during the quiescent state. In this report I will first make an overview of the TeV extragalactic sky, showing that the quasi totality of the VHE gamma-ray emitters are active galactic nuclei (AGN). I will then introduce the current paradigm explaining the different flavours of jetted AGNs with a particular emphasis on the blazar sequence, recently revised. After this introduction I will start the detailed report on recent highlight results from the current IACTs, with three sections dedicated to different TeV emitters, and a last section focused on cosmology and propagation studies with VHE gamma rays. I will conclude with a short summary and a list of future perspectives of this exciting and very fecund research field.
:::
# An overview of the TeV extragalactic sky {#sec:TeVsky}
A list of TeV-emitting extragalactic sources detected by current generation of IACTs is reported in Table [1](#tab:src){reference-type="ref" reference="tab:src"} and plotted in Figure [\[fig:TeVsky\]](#fig:TeVsky){reference-type="ref" reference="fig:TeVsky"}. The main features characterizing the detected sources are:
- The large majority of the TeV extragalactic emitters are classified as high synchrotron peaked BL Lac objects (see next section for details);
- The only non-AGN sources are two nearby starburst galaxies: M 82 and NGC 253 (orange filled dots in the Figure);
- The redshift is limited below 1, with the large majority of the sources featuring a redshift below 0.5;
- A number of detected blazars have no redshift measurement. This is a quite typical feature of BL Lac objects, that may have no emission/absorption lines and whose optical non-thermal continuum is so strong to prevent the detection of the host galaxy.
# Jetted AGNs and the Blazar Sequence {#sec:AGNmodel}
An AGN is defined as a compact region at the center of a galaxy showing a luminosity much higher than normal, mostly of non-stellar origin. A plethora of different AGN categories have been proposed in literature starting from the early seventies, according to some spectral features such as the characteristics of emission and absorption lines in the spectrum or the radio luminosity. The latter criterion separates the sample into the two main classes of radio-loud and radio-quiet AGNs. In 1995, after several years of astronomical discoveries, Urry and Padovani proposed the innovative idea that different features seen in the spectra of radio-loud AGNs could be related to the different orientation of the same, complex, object lying in the nucleus of the active galaxies. Since then, radio loud AGNs have been described as supermassive black holes accreting material and ejecting part of it through two collimated jets of particles accelerated to ultra-relativistic energies (in radio quiet objects the structure is similar but there are no jets). Very recently, in order to include a sample of radio-faint objects showing many characteristics in common with radio-loud AGNs, the new definition of *jetted AGNs* instead of radio-loud AGNs has been proposed, and will be used in what follows. Jetted AGNs usually show a double-peaked SED covering almost all the electromagnetic spectrum, from radio to gamma-ray frequencies. The first peak of the SED is due to synchrotron radiation by accelerated electrons in the jet, and the second is most likely related to inverse Compton emission. Additional components in the SED related to the light emitted by the accretion disc or a hot corona may be present. For a recent review see. The two main classes of jetted AGNs are:
Radiogalaxies
: : when the viewing angle connecting the observer with the jet direction is relatively large (\(>\)`<!-- -->`{=html}20\(^{\circ}\)). In radio, these objecs usually show two thin radio structures extending up to several kpc and identified as the emission from the two jets.
Blazars
: : when the viewing angle connecting the observer with the jet direction is relatively small (\(<\) 20\(^{\circ}\)). Due to ultra-relativistic velocities of the particles in the jet, the jet emission is stronlgy boosted and dominates the overall luminosity. Depending on the presence of emission lines in the spectrum, blazars are further divided into Flat Spectrum Radio Quasars (FSRQs) and BL Lac objects, with the former characterized by strong broad emission lines in the optical spectrum.
A very interesting feature of blazars SED is the anticorrelation between their bolometric luminosity and the position of the synchrotron peak, the so-called blazar sequence. A recent study performed by Ghisellini explores the same sequence in a subsample of \(\sim\)`<!-- -->`{=html}750 blazars detected by *Fermi*-LAT, and confirms that also the second peak of the SED follows the same trend (even if less pronounced). FSRQs show a very bright spectrum with the SED peaks located mainly towards low energies while BL Lacs feature a lower luminosity and the peaks shifted at higher energies. The subclasses of low/intermediate/and high synchrotron peaked BL Lac objects (LBL, IBL, and HBL respectively) seem to reflect this behaviour. Looking at this sequence, it is not surprising that the large majority of blazars detected by IACTs are HBLs: blazars with a synchrotron and a high energy peak shifted towards the highest energies, while FSRQs are usually detected at VHE only during flaring episodes. After almost fifty years of jetted-AGN observations, we have now many experimental evidences supporting a general picture, but also many unknowns. These are: the size and location of the emitting region, the role of hadrons in the jet, and the acceleration mechanism at work (shocks or other processes). Current IACTs are optimizing the observation time in order to discover new sources of VHE radiation and study in detail the emission from known sources. In both cases, the collaboration with telescopes observing in other bands is crucial. New discoveries are in fact very often triggered by an optical or gamma-ray high-state of a good candidate, while detailed studies are carried out in a multi-wavelength approach.
## The Gamma-ray horizon
The limited redshift distribution that characterizes the TeV-detected blazars, if compared for example to the blazars detected at lower frequencies by *Fermi*-LAT satellite, can only in part be attributed to the extreme energetics needed to produce TeV gamma rays. The other reason is related to the interaction of VHE photons with the optical and infrared photons filling the Universe, the so-called extragalactic background light (EBL). EBL is the sum of the light emitted by the stars and reprocessed by the dust since the birth of the first stars. When a TeV gamma ray is produced, it may interact with EBL photons and create an electron/positron pair. This process depends on the energy of the photon and on the distance of the emitter: in closeby sources only the most energetic photons are affected (above few TeV), while at larger redshift (\(>\)`<!-- -->`{=html}0.3) this process is already effective at a few hundreds GeV. For sources located at redshift 1, 100 GeV photons are already significantly absorbed. Different EBL models have been developed, which try to describe the EBL energy density and its evolution with cosmic time. Last generation of models, e.g., are pretty in agreement and seem to suggest that we are close to resolve the quasi totality of the sources emitting optical radiation.
# Gamma rays from misaligned blazars {#sec:rgal}
In view of the unified model presented in previous section, radiogalaxies can be considered as *misaligned blazars*. Thanks to the large viewing angle, they offer the unique opportunity to localize the emitting region of blazars. Four of these objects have been up to now detected up to the extreme energies, namely PKS 0625--35, M 87, NGC 1275, and Cen A. Additionally, IC 310 can be included in this family due to its uncertain nature. Interesting results from Cen A and IC 310 were recently published and I will shortly describe them in the following subsections. The interested reader is addressed to and references therein for an exhaustive description.
## Cen A
Centaurus A is the closest and best studied radio galaxy. Its morphology was observed in detail and it is quite complex, showing jets at different scales (sub-pc to kpc), radio lobes, and an extended emission in many bands, including gamma rays. Usually the broadband SED from the core of radio galaxies is well described by a synchrotron self-Compton (SSC) model, with the low energy emission due to synchrotron radiation emitted by relativistic electrons, and the high energy component due to the very same electrons upscattering the synchrotron photons via inverse Compton mechanism. The detection of an unexpected hardening in the high energy spectrum from 7.5 years of *Fermi*-LAT data in the core of Cen A was recently published in. Left panel of Figure [\[fig:MisalignedBzb\]](#fig:MisalignedBzb){reference-type="ref" reference="fig:MisalignedBzb"}, from, shows the spectrum measured with *Fermi*-LAT (black dots) together with the long-term VHE spectrum measured with H.E.S.S. (red triangles). Interestinlgy, below few GeV the spectral slope is compatible with the slope found at the highest energies, which are however shifted toward a higher flux density. The spectral hardening detected in *Fermi* data well connects these two regimes and strongly supports the hypothesis of an *additional emission component in the spectrum of Cen A* extending at VHE. According to the authors, this component could be due to a spike in the dark matter profile, to a population of millisecond pulsars, or to an additional hadronic component.
## IC 310
Another uncommon feature that has been recently detected in a misaligned blazar is the extremely short timescale variability seen in IC 310, a known TeV emitter located in the Perseus cluster of galaxies. The MAGIC telescopes observed IC 310 in November 12-13, 2012. The light curve measured above 300 GeV, reported in right panel of Figure [\[fig:MisalignedBzb\]](#fig:MisalignedBzb){reference-type="ref" reference="fig:MisalignedBzb"}, revealed an incredibly fast variability with a flux doubling time shorter than 5 minutes. This measure allowed a very constraining measurement of the size of the emitting region. Considering that causality forces this size *s* to be smaller than \(c\,\delta t\), \(s\) results smaller than 20% of the gravitational radius of the central black hole. Also in this case, this experimental evidence marks a break of the standard paradigm of emission from the nucleus of a jetted active galaxy. According to the standard SSC model, in fact, the size of the emitting region cannot be smaller than the jet width, as in IC 310. The most natural explanation of this finding is the occurrence of vacuum gap-type particle acceleration on sub-horizon scales, a mechanism often invoked to explain pulsar emission.
# Gamma rays from FSRQs {#sec:fsrq}
FSRQs are the majority of extragalactic sources detected in the high energy range by *Fermi*-LAT, while they constitute just a small fraction of the TeV-detected sources. As mentioned above, this evidence can be fully explained with the blazar sequence: FSRQs are very bright objects whose SED high-energy peak is located in the sub-GeV range. Moreover, the typical redshift of FSRQs is well above 1, making them difficult targets for IACTs due to the interaction with EBL. In the last few years, VHE observations have challenged the standard emission model proposed for these objects, which explained the observed non-thermal continuum as syncrotron radiation (at low energies) and inverse Compton emission on external photon fields (also called *external Compton*, at higher energies). In FSRQs, the presence of an internal region close to the central object characterized by a dense photon field (the *broad line region, BLR*), implies that if VHE radiation is emitted within this region, it must be self-absorbed. The issue arised when an unexpected short variability timescale was observed at VHE from the FSRQs PKS 1222+216. Causality forces the emitting region to be very compact, meaning close to the black hole if we assume the canonical shock-in-the jet model. But no VHE emission should escape this region due to self absorption, as mentioned above. Seven FSRQs have been up to now detected at the highest energies. Two of them are located at surprisingly high redshift: PKS 1441+25 at \(z = 0.939\) and B0218+357 at \(z = 0.944\), and will be briefly outlined.
## PKS 1441+25
A significant VHE signal from PKS 1441+25 was detected in the dataset collected with the MAGIC telescopes the second half of April 2015. The observations were triggered by a high emission state seen with *Fermi*-LAT. VERITAS also observed the source and detected a significant signal. A dense multi-wavelength campaign on the source was carried on during this episode. The large distance of PKS 1441+25 implies a very strong absorption of the VHE spectrum already above 100 GeV. Very remarkably, once corrected for the EBL aborption the spectrum connects smoothly with the simultaneous spectrum measured by *Fermi*-LAT and the overall SED is nicely fitted with a SSC model with an additional component due to external Compton of infra-red photons coming from a dusty torus sourrunding the accretion disk, see left plot in Figure [\[fig:FSRQs\]](#fig:FSRQs){reference-type="ref" reference="fig:FSRQs"}. The variability timescale is constrained to few days, and is in agreement with the shock-in-the-jet model once assumed that the emitting region lies outside the BLR.
## B0218+357
Two main features characterize the FSRQ B0218+357: its large redshift, making this source the most distant TeV emitter known to date, and the fact that this source is a gravitationally lensed blazar. The lens is the spiral galaxy B0218+357G located at a redshift of \(z~\sim 0.68\). The radio image clearly shows two components, with a delay between the two images measured during flaring episodes of 10-12 days. This delay was also seen in 2012 at high energies by *Fermi*-LAT during a series of outbursts. The successive strong flare from the source was registered by *Fermi* in 2014, and triggered observations with the MAGIC telescopes. However MAGIC could point the source only during the expected arrival time of the delayed component of the emission. A significant signal was detected by MAGIC below 200 GeV. Figure [\[fig:FSRQs_0218\]](#fig:FSRQs_0218){reference-type="ref" reference="fig:FSRQs_0218"} shows the VHE and high-energy gamma rays, X-ray, and optical light curve of this event. Interestingly, two distinct emitting region in the jet are necessary to explain the observed SED, which is well reproduced in the framework of a two zones external Compton scenario, as detailed in .
# Gamma rays from BL Lac objects {#sec:blac}
BL Lacs form the most numerous class of extragalactic objects seen at VHE gamma rays. In the last five years, IACTs discovered more than 10 new sources of this class, mainly thanks to high-energy gamma-, X-ray and optical alerts (e.g. the very recent detection of the blazar OJ 287 reported by the VERITAS team in the ATel \#10051). Moreover, a number of known emitters are monitored regularly with quasi simultanoeus data collected in different bands. These multi-wavelength campaigns are aimed at characterizing the physical properties of the jet. An example of monitoring during a quiescent state of Mkn 421 is presented in , where MAGIC and VERITAS data are combined for the first time with NuSTAR (hard X-ray band) and many other data from radio to gamma rays to build and model a very detailed multi-epoch SED. Known sources are often use to calibrate instruments, like the case of PG 1553+113.
## PG 1553+113
PG 1553+113 is one of the few sources that can be seen by all the three main IACTs currently in operation. It is a BL Lac with unknown redshift, most likely lying between 0.43 and 0.58. Currently PG 1553+113 is the object of a deep, regular multi-wavelength campaign at several bands including VHE gamma rays. This campaign followed the claim of a quasi-periodicity in the gamma ray and optical lightcurves seen by the *Fermi*-LAT team. Together with PKS 2155-304, PG 1553+113 was used by the H.E.S.S. team to study in detail the performances of the CT5 telescope in mono mode. Figure [\[fig:bllac\]](#fig:bllac){reference-type="ref" reference="fig:bllac"} shows the gamma-ray SED of the source measured with simultanoeus CT5 and *Fermi*-LAT data. The CT5 spectrum is presented with significant data down to 110 GeV and connects smoothly with lower energy data measured by the LAT instrument. No significant curvature was seen in the spectrum of PG 1553+113.
# Cosmology and Propagation with VHE gamma rays
Several studies conducted in the last years have demonstrated that VHE gamma rays from cosmological sources are a very useful tool for studying cosmology and photon propagation. In particular VHE gamma rays can be used to test Lorentz invariance violation models, probe the intergalactic magnetic field, test EBL models and the hypothesis of photon oscillation in axion-like particles. I will shortly report on the last two subjects. For a recent review see and references therein.
## EBL
The effect on blazars spectra due to the interaction of VHE photons with EBL photons can be used to probe, and in some cases exclude, the EBL models (e.g. ). This can be done once an intrinsic shape for the blazar spectra is assumed. First observations of distant blazars, such the FSRQ 3C 279 located at redshift 0.536 and discovered at VHE by MAGIC in 2007, provided constraints on some EBL models in use ten years ago. Once corrected for the absoption in fact, the instrinsic spectrum showed a significant pile up at the highest energies that is very difficult to reconcile with our knowledge of blazar emission. Current models are converging on a similar shape for the EBL. Recent detections of blazars up to redshift \(\sim\) 1 allowed to probe for the first time new parts of the EBL spectrum (those at the highest energies, in the optical range). Figure [\[fig:cosmo\]](#fig:cosmo){reference-type="ref" reference="fig:cosmo"} shows the limits provided by VERITAS measurement of the spectrum of PKS 1441+25, already mentioned above. The upper limits set are very close to EBL firm lower limits given by direct galaxy counts and are remarkably in agreement with EBL models. A multi-source approach have been proposed in several studies. In particular Biteau and Williams used a sample of 106 spectra from 36 different objects and were able to build their own EBL spectrum from mid-IR to far IR. This spectrum is in good agreement with current models and upper/lower limits.
## Axion-like particle hypothesis
The long distances travelled by gamma rays into the interstellar space offer the possibility of testing the hypothesis of photons oscillation. A very popular proposal is that VHE gamma rays may oscillate into the so-called axion-like particles (ALP) with a probability related to the external magnetic field, the distance of the emitter, and the \(\gamma\)-ALP coupling constant (see for example ). Hence, this oscillation (including the ALP/\(\gamma\)-ray reconversion) may happen in the vicinity of the source, in the intergalactic space, or in our galaxy. The hypothetical effect of these oscillations is a distorsion of the VHE spectra of an AGN and a partial reduction of the EBL suppression effect, since ALPs do not interact with EBL photons. Possible signatures may be wiggles in the VHE spectra of blazars or the observation of appartent high-opacity photons. In the last years different studies led to different, in some cases contradictory, results. The analysis of the 106 spectra of blazars by Biteau and Williams resulted into the non detection of any deviation from the expected flux once corrected by EBL effect. This is in contradiction with the claim done in, where an evidence for such a deviation is instead found. Also in this case, good quality spectra of AGNs, in particular those at large distances, would be the smoking gun to probe (or discard) this hypothesis.
# Summary and Conclusions
In summary, VHE observations of jetted AGNs provided by IACTs are important for different aspects. The lightcurves sampled with the good time resolution of IACTs are a very powerful tool to constrain the size of the emitting region. Spectral measurements allow to probe the highest energies reached in the jet and test the emission mechanism at work. Moreover, VHE photons, interacting with the EBL, can be used to test EBL models or the hypothesis of oscillation of photons in axion-like particles. The last few years have seen a number of important results covering many aspects of the extragalactic VHE gamma ray sky. While some results are in pretty good agreement with current models (such as the EBL absoption effect), some others, in particular those related to fast variability, strongly suggest that the standard paradigm of radiogalaxy/blazar emission should be revised. Further measurements of timing and spectral properties of the broadband emission from TeV AGNs are therefore mandatory. This is the main challenge for the IACTs of the next decade, and in particular for the CTA, which is expected to increase the number of known sources by at least one order of magnitude. | {'timestamp': '2017-07-06T02:03:08', 'yymm': '1706', 'arxiv_id': '1706.01670', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01670'} |
# Introduction {#sec:1}
Diffusion MRI is a powerful non-invasive imaging technique widely used to explore white matter in the human brain. Diffusion Tensor Imaging (DTI) is used to reconstruct a tensor field from diffusion weighted images (DWIs). High Angular Resolution Diffusion Imaging , which makes no assumption of a 3D Gaussian distribution of the diffusion propagator, is used to reconstruct a general function field from DWIs, (e.g., an Orientation Distribution Function (ODF) or Ensemble Average Propagator (EAP) field). Both the ODF and the EAP fields with a given radius are spherical function fields. Exploring microstructural information from the reconstructed tensor field or spherical function field is of interest in many biological and clinical application areas, which makes diffusion MRI a powerful means to study white matter. For example, in an voxel exhibiting anisotropic diffusion, local peaks of the reconstructed spherical function or the principal eigenvector of the reconstructed 2nd-order diffusion tensor normally prescribe the fiber directions in that voxel. Some scalar indices have been proposed to be estimated voxel-wise from tensors/ODFs/EAPs. For DTI, well-established tensor scalar indices, including the mean diffusivity and Fractional Anisotropy (FA), are widely used as biologically meaningful descriptors . proposed two sets of scalar indices (three scalar indices per set), which are orthogonal in terms of tensor changes and the Euclidean inner product. For High Angular Resolution Diffusion Imaging (HARDI), the generalized FA , Orientation Dispersion index (OD) , return-to-origin probability , and mean-squared displacement were all proposed for ODFs and EAPs. These indices indicate some information inside a voxel, but cannot describe local geometric or topological information, including fiber crossing, fanning, bending, and twisting, in a local spatial neighborhood. Some previous works have extracted local geometric information by considering the local spatial change of tensor fields or ODF fields. demonstrated that the norm of the spatial gradient of the tensor's isotropic and anisotropic parts can detect boundaries between white matter, CSF, and gray matter. proposed tangents of scalar invariants and rotation tangents, which are 2nd-order tensors, and also proposed projecting the 3rd-order spatial gradient tensor onto these 2nd-order tangents to obtain the spatial direction with the largest change of scalar indices or rotation of tensors. Based on the rotation tangents of tensors, proposed fiber curving and fiber dispersion indices. proposed a sheet probability index to quantify the local sheet structure by using spatial changes of ODF peaks. proposed spatial and spherical smoothing to enhance an ODF field in a PDE framework, preserving crossing structures. considered spatial coherence in ODF estimation. The terms "splay", "bend" and "twist" have been used to qualitatively describe complex local white matter structural configuration in literature of diffusion MRI for about 20 years . However, to our knowledge, there is no existing work that *quantitatively* describes the degree of local orientational change of white matter, including splay, bend, and twist, from general ODF fields in dMRI, although the fiber curving and dispersion indices in can be seen to quantify "splay" and "bend" for a tensor field in DTI. discussed the initial idea to study the torsion and curvature of a fiber tract by using the Frenet frame [^1] along the tract. Torsion and curvature from the Frenet frame were later used in diffusion data analysis in . used the Frenet frame as a prior in the relaxation labeling algorithm to regularize the data and estimate ODFs in voxels. These works on the Frenet frame studied geometric information along a single tract. However, tractography is known to be sensitive to a large number of parameters, and any flaws in the reconstructed tracts due to noise or parameter selection will inevitably be reflected in the geometric information that is extracted subsequently. proposed moving frames determined by the geometry of cardiac data, and calculated Maurer-Cartan connections. However this method is not applicable to general diffusion MRI data, and does not consider the sign ambiguity in the frame. There exist some connections between diffusion MRI data analysis and liquid crystals. Orientational order parameter is well-established to describe the degree of alignment in liquid crystals . calculated the order parameter map by estimating variance of microscopic diffusion parameters from the contrast between diffusion signals measured by directional and isotropic diffusion encoding. However, it cannot be used for general DTI and HARDI data. used a diffusion tensor method to estimate the director orientations of a lyotropic liquid crystal as a spatially resolved field of Saupe order tensors. In this paper, inspired by orientation and distortion analyses applied to liquid crystals, we propose a unified framework, called Director Field Analysis (DFA), to study the local geometric information of white matter from the reconstructed spherical function field. DFA works both for tensor fields obtained from DTI and for spherical function fields from HARDI. At the voxel level, 1) the Orientational Order index (OO) and Orientational Dispersion index (OD) are defined for the spherical function in a voxel with a given axis (e.g., the ODF with its principal direction); and 2) the principal direction is extracted from the spherical function in such a voxel exhibiting anisotropic diffusion. At a local neighborhood level, 1) an orthogonal coordinate frame is defined for each voxel with anisotropic diffusion, where the first axis is the extracted principal direction; 2) OO is defined for spherical functions in a local neighborhood with the given principal direction; and 3) three distortion indices (splay, bend, twist) and a total distortion index are defined based on the spatial directional derivatives of the principal direction. An overview of the DFA pipeline for a spherical function field is shown in Fig. [\[fig:DFA\]](#fig:DFA){reference-type="ref" reference="fig:DFA"}. This paper is organized as follows. Section [2](#sec:method_tensor){reference-type="ref" reference="sec:method_tensor"} provides a unified overview of existing works on tensor field analysis for exploring local geometric information , which is also a motivation for the proposed DFA. Section [\[sec:method\]](#sec:method){reference-type="ref" reference="sec:method"} proposes the DFA framework that works for both diffusion tensor fields and ODF fields. Section [\[sec:experiments\]](#sec:experiments){reference-type="ref" reference="sec:experiments"} demonstrates some results of synthetic and real data experiments by using DFA. Section [\[sec:discussion\]](#sec:discussion){reference-type="ref" reference="sec:discussion"} discusses some issues on implementing DFA.
# Tensor Field Analysis {#sec:method_tensor}
This section provides an overview of existing works exploring the local geometric features of a 2nd-order diffusion tensor field by using the spatial gradient of tensors in a unified framework. It also proposes a new 4th-order structure tensor applied to 2nd-order tensor data that generalizes the conventional 2nd-order structure tensor applied to scalar fields. **The 2nd-order diffusion tensor field**. In a diffusion tensor field denoted as \(\BM{D}\), there is a diffusion tensor \(\BM{D}(\mathbi{x})\) at the voxel \(\mathbi{x}\), where \(\BM{D}(\mathbi{x}) \in S_+^3\), and \(S_+^3\) is the set of \(3\times 3\) symmetric positive definite matrices. The diffusion tensor is symmetric with six unique (i.e., independent) elements. **The 3rd-order spatial gradient of the diffusion tensor**. For a tensor field denoted as \(\BM{D}\) with elements \([D_{ij}(\mathbi{x})]\) at the voxel \(\mathbi{x}\), its spatial gradient at voxel \(\mathbi{x}\), denoted as \(\nabla_\mathbi{x} \BM{D}(\mathbi{x})\), is a 3rd-order tensor with elements \([D_{ij,k}(\mathbi{x})=\frac{\partial D_{ij}}{\partial x_k}(\mathbi{x})]\), where \(i,j,k\in \{1,2,3\}\). Since the diffusion tensor is symmetric with six unique elements, the 3rd-order spatial gradient has 18 unique elements. **Mapping the 3rd-order spatial gradient to a vector**. Let \(\BM{W}=[W_{ij}]\) be a designed 2nd-order weighting tensor in tensor space, then the tensor inner product
[^1]: <https://en.wikipedia.org/wiki/Frenet-Serret_formulas> | {'timestamp': '2017-10-17T02:05:28', 'yymm': '1706', 'arxiv_id': '1706.01862', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01862'} |
null | null |
# Introduction
Let \((X,\rho)\) be a metric space. For a compact subset \(K\) of \(X\), \(b_{(X,\rho)}(K)\) denotes the *Borsuk number* of \(K\), that is, the minimal number of parts of \(K\) of smaller diameter necessary to partition \(K\). The *Borsuk number* of \((X,\rho)\) is defined to be \(B(X,\rho)=\max _{K}b_{X,\rho}(K)\). If \(\Omega(\rho)\) denotes the set of metrics on \(X\) equivalent to \(\rho\), then the *topological Borsuk number* of \((X,\rho)\) is defined by \(B(X)=\min_{\tau\in\Omega(\rho)}B(X,\tau)\). The *topological Borsuk problem* for \(\mathbb R^d\) is to estimate the *topological Borsuk number* of \(\mathbb R^d\), \(B(\mathbb{R}^n)\) for the Euclidean metric (see ); one wants to know if \(B(\mathbb{R}^n)\geq n+1\). Y. Soibelman proved in that \(B(\mathbb{R}^2)=3\), but it is not known if \(B(\mathbb{R}^3)\geq 4\). A nonempty subset \(S\) of a metric space \((X,\rho)\) is called *equilateral (with respect to \(\rho\))*, if there is some positive real number \(c\) such that \(\rho(x,y)=c\) for all \(x,y\in S\), \(x\neq y\). In order to prove that \(B(\mathbb R^d)\geq m\), it suffices to show that for any metric \(\rho\) defined on \(\mathbb R^d\) equivalent to the Euclidean metric, there exists some equilateral subset \(S\subseteq (\mathbb R^d,\rho)\) of size \(m\) (see ). The following result is due to C. M Petty ():
Theorem [\[petty\]](#petty){reference-type="ref" reference="petty"} implies that for any metric \(\rho\) on \(\mathbb R^3\) induced by a norm we have \(B(\mathbb R^2, \rho)\geq 4\). In fact, restrict \(\rho\) to a 2-plane \(X\) in \(\mathbb R^3\) and apply the result of Soibelman that the Borsuk number of \(\mathbb R^2\) is 3 to obtain a \(\rho\)-equilateral set in \(X\) of size 3. This equilateral set in \(X\) is also an equilateral set of size 3 in \((\mathbb R^3,\rho)\), so by theorem [\[petty\]](#petty){reference-type="ref" reference="petty"}, it extends to a \(\rho\)-equilateral set of size 4. Thus, \(B(\mathbb R^4, \rho)\geq 4\). This gives us a partial answer to the Borsuk problem for \(\mathbb R^3\). An *\(m\)-gon* in a metric space \((X,\rho)\) is an \(m\)-cycle graph \(G\) whose vertices are \(m\) distinct elements in \(X\). If \(G\) has vertices \(x_0,x_1,\ldots,x_{m-1}\) and edges \(\{x_i,x_{i+1}\}\) for \(i=0,1,\ldots,m-1\) (where \(x_{m}=x_0\)), we say that \(G\) is an *equilateral \(m\)-gon* in \((X,\rho)\), or a *equilateral \(m\)-gon with respect to \(\rho\)*, if there is some constant \(c>0\) such that \(\rho(x_i,x_{i+1})=c\) for all \(i=0,1,\ldots,m-1\). Given \(r\) metrics on \(X\), \(\rho_1,\ldots,\rho_r\), we say that \(G\) is an *\(r\)-equilateral \(m\)-gon* (with respect to \(\rho_1,\ldots,\rho_r\)) if it is equilateral for each metric \(\rho_i\). Let \(\mathbb{Z}/m=\langle \omega|\ \omega^m=1\rangle\). Given an \(m\)-gon \(G\) with vertices \(x_0,x_1,\ldots,x_{m-1}\) and edges \(\{x_i,x_{i+1}\}\) for \(i=0,1,\ldots,m-1\) (where \(x_{m}=x_0\)), \(\mathbb{Z}/m\) acts naturally on the set of vertices of \(G\) by \(\omega\cdot x_i=x_{i+1}\), and this induces an action of \(\mathbb{Z}/m\) on the set of edges of \(G\): \(\omega\cdot \{x_i,x_{i+1}\}=\{\omega\cdot x_{i},\omega\cdot x_{i+1}\}\). If \(K_m\) denotes the complete graph on vertices \(x_0,\ldots,x_{m-1}\), then the action of \(\mathbb{Z}/m\) on \(x_0,x_1,\ldots, x_{m-1}\) induces an action of \(\mathbb Z/m\) on the edges of \(K_m\). The set of edges of \(K_m\) decomposes into a union of disjoint orbits; some of these orbits represent \(m\)-gons, some other are \(n\)-gons for some \(n<m\) and some other orbits are not even \(n\)-gons for any \(n\) (see figure [\[figmgons\]](#figmgons){reference-type="ref" reference="figmgons"} (a)). If \(m=p\) is an odd prime number, then each orbit of \(\mathbb{Z}/p\) acting on the set of edges of \(K_p\) is actually a \(p\)-gon, and there are \((p-1)/2\) of them (see figure [\[figmgons\]](#figmgons){reference-type="ref" reference="figmgons"} (b)). If the subscripts of the \(x_i\)'s are thought as the elements of the finite field \(\mathbb{F}_p\), then these \((p-2)/2\) \(p\)-gons, that we will denote by \(C_1,C_2,\ldots, C_{(p-1)/2}\), can be described as follows: \(C_t\) has edges \(\{x_{0}, x_{t}\}, \{x_{t}, x_{2t}\}, \ldots, \{x_{(p-1)t}, x_{0t}\}\). Our first main result is the following:
As a consequence, if we put \(t_1=t_2=\cdots=t_r\) in theorem [\[equilateral\]](#equilateral){reference-type="ref" reference="equilateral"}, then we have that there exists an \(r\)-equilateral \(p\)-gon in \(\mathbb R^d\) with respect to \(\rho_1,\rho_2,\ldots,\rho_r\). To state our second main result we first introduce some terminology. Let \(p\) be an odd prime and \(d\geq 1\). We call the pair \((d,p)\) *admissible* if given any continuous injective map \(f:S^{d}\to \mathbb{R}^{d+1}\) and \((p-1)/2\) metrics \(\rho_1,\ldots,\rho_{(p-1)/2}\) on \(\mathbb{R}^{d+1}\), each metric equivalent to the Euclidean metric of \(\mathbb{R}^{d+1}\), then there can be found \(p\) different points in the image of \(f\), \(f(x_0), \ldots,f(x_{p-1})\), such that \[\rho_j(f(x_0), f(x_{j})) =\rho_j(f(x_{j}), f(x_{2j}))=\cdots=\rho_j(f(x_{(p-1)j}), f(x_0)),\] for all \(j=1,2,\ldots,(p-1)/2\). In other words, the \((p-1)/2\) \(p\)-gons determined by \(f(x_0), \ldots,f(x_{p-1})\), \(C_1,C_2,\ldots, C_{(p-1)/2}\) satisfy that each \(C_i\) is equilateral with respect to \(\rho_i\), \(i=1,2,\ldots, (p-1)/2\). Now we can state our result as follows:
Both results, Theorem [\[equilateral\]](#equilateral){reference-type="ref" reference="equilateral"} and Theorem [\[pgonsdeformed\]](#pgonsdeformed){reference-type="ref" reference="pgonsdeformed"} are proved by means of topological methods: we use the *configuration space/test map scheme* and the *Fadell-Husseini index with coefficients in \(\mathbb F_p\)* of the involved spaces to solve our problems.
# Fadell-Husseini index
Let \(G\) be a finite group and \(R\) a commutative ring with unit. The Fadell-Husseini index with coefficients in \(R\) assigns to each \(G\)-space \(X\) an ideal \(Ind_{G,R}X\) of the ring \(H^{\ast}(BG;R)\), the cohomology ring of the classifying space \(BG\) with coefficients in \(R\). The important property of the Fadell-Husseini index is that it gives a necessary condition for the existence of equivariant maps between to \(G\)-spaces: if \(X\) and \(Y\) are \(G\)-spaces and there is some \(G\)-equivariant map \(f:X\to Y\), then \(Ind_{G,R}Y\subseteq Ind_{G,R}X\). For definitions, basic properties and some calculations see. Here, we want to mention the properties and calculations of the Fadell-Husseini index we will use in the proofs of our results. For a finite group \(G\) we have the associated universal \(G\)-bundle \(EG\to BG\); if \(X\) is a \(G\)-space \(X\), then there is the associated *Borel construction* \(X_G=X\times_GEG\) and a map \(q_X:X_G\to BG\) induced by the unique map \(X\to \ast\). The index \(Ind_{G,R}X\) is defined as the kernel of the induced map in cohomology \(q_X^{\ast}:H^{\ast}(BG;R)\to H^{\ast}(X_G;R)\). The *axioms of an ideal valued index theory* are satisfied by Fadell-Husseini index:\
**Monotonicity:** If there is some \(G\)-equivariant map \(f:X\to Y\), for \(G\)-spaces \(X\) and \(Y\), then \[Ind_G Y\subseteq Ind_G X.\] **Additivity:** If \(\{X=X_1\cup X_2,X_1,X_2\}\) is excisive, where \(X_1\) and \(X_2\) are \(G\)-invariant subspaces of \(X\), then \[Ind_G\ X_1\cdot Ind_G\ X_2\subseteq Ind\ X.\] **Continuity:** If \(A\) is a closed \(G\)-invariant subspace of \(X\), then there is a \(G\)-invariant neighborhood \(U\) of \(A\) in \(X\) such that \[Ind_G\ A=Ind_G\ U.\] **Index theorem:** If \(f:X\to Y\) is \(G\)-equivariant, \(B\subseteq Y\) a closed \(G\)-invariant subspace and \(A:=f^{-1}(B)\), then \[Ind_G\ A\cdot Ind_G\ (Y\setminus B)\subseteq Ind_G\ X.\] For instance, if \(X\) is a \(G\)-space and \(Y\) an \(H\)-space, then \(X\times Y\) is a \(G\times H\)-space; the map \(q_{X\times Y}:(X\times Y)_{G\times H}\to B(G\times H)\) can be identified with the map \(q_X\times q_Y:X_G\times Y_H\to BG\times BH\) and the map \(q_{X\times Y}^{\ast}\) can be identified (under Künneth isomorphisms) with \[q_X^{\ast}\otimes q_Y^{\ast}: H^{\ast}(BG ;\mathbb{K})\otimes H^{\ast}(BH; \mathbb{K})\to H_G^{\ast}(X; \mathbb{K})\otimes H_H^{\ast}(Y; \mathbb{K}),\] where \(\mathbb{K}\) is a field. The kernel of this map can be expressed as follows: \[\ker q_X^{\ast}\otimes q_Y^{\ast}=\ker q_X^{\ast}\otimes H^{\ast}(BH; \mathbb{K})+H^{\ast}(BG; \mathbb{K})\otimes \ker q_Y^{\ast},\] that is, \[Ind_{G\times H,\mathbb{K}} X\times Y=Ind_{G,\mathbb{K}}\otimes H^{\ast}(BH; \mathbb{K})+H^{\ast}(BG; \mathbb{K})\otimes Ind_{H,\mathbb{K}}Y.\] As particular cases, if \(Y=pt\), then \[Ind_{G\times H,\mathbb{K}}X=Ind_{G,\mathbb{K}}X\otimes H^{\ast}(BH; \mathbb{K});\] and if \(H^{\ast}(BG; \mathbb{K})\cong \mathbb{K}[x_1,\ldots, x_n], H^{\ast}(BH; \mathbb{K})\cong \mathbb{K}[y_1,\ldots, y_m]\), \(Ind_{G,\mathbb{K}}X=\langle f_1,\ldots, f_r\rangle \text{ and } Ind_{H,\mathbb{K}}Y=\langle g_1,\ldots, g_s\rangle,\) then \[Ind_{G\times H,\mathbb{K}}X\times Y=\langle f_1,\ldots,f_r,g_1,\ldots, g_s\rangle \subseteq \mathbb{K}[x_1,\ldots,x_n,y_1,\ldots, y_m].\] These results on the index of products can be consulted in. **Spectral Sequence.** The map \(q_X:X_G\to BG\) is a fibration with fiber \(X\). There is a spectral sequence \(\{E_r^{\ast,\ast},d_r\}\) with \(E_2^{p,q}=H^p(BG;\mathcal{H}^q(X;R))\) where \(\mathcal{H}^q(X;R)\) is a system of local coefficients, this local coefficient system is determined by the action of \(G=\pi_1(BG)\) on \(H^{\ast}(X;R)\). Therefore, the \(E_2\)-page of the spectral sequence is interpreted as the cohomology of the group \(G\) with coefficients in the \(G\)-module \(H^{\ast}(X,R)\) (see for example ), \[E_2^{p,q}=H^p(G;H^q(X;R)).\] The map \(q_X^{\ast}:H^{\ast}(BG,R)\to H^{\ast}(X_G,R)\) can be represented as the composition (known as the *edge homomorphism*) \[H^{\ast}(BG;R)\to E_2^{\ast,0}\to E_3^{\ast,0}\to E_4^{\ast,0}\to\cdots\to E_{\infty}^{\ast,0}\subseteq H^{\ast}(X_G;R).\]\
**Spheres, \(E_nG\)-spaces and the configuration space \(F(\mathbb R^d,p)\).** A free \(G\)-space \(X\) is an *\(E_nG\)-space* if it is an \((n-1)\)-connected, \(n\)-dimensional CW-complex equipped with a \(G\)-invariant CW-structure. \(E_nG\)-spaces exist; in fact, the join \(G^{\ast(n+1)}\) of \(n+1\) copies of the group \(G\), where \(G\) is regarded as a 0-dimensional simplicial complex, is an \(E_nG\)-space. The following properties of the Fadell-Husseini index can be found in.
The *configuration space* of \(n\) distinct points in the topological space \(X\) is the space \[F(X,n)=\{(x_1,x_2,\ldots,x_n)\in X^n: x_i\neq x_j\text{ for all} i\neq j\}.\] The symmetric group \(S_n\) acts on \(F(X,n)\) by permuting coordinates; the cyclic group \(\mathbb Z/n\) seen as the subgroup of \(S_n\) generated by the permutation \((12\cdots n)\), acts on \(F(X,n)\) by restricting the action of \(S_n\). The computation of \(Ind_{\mathbb Z/p,\mathbb F_p}F(\mathbb R^d)\), where \(p\) is an odd prime number and \(d>1\) is made in (Theorem 6.1): \[\label{indexconfiguration} Ind_{\mathbb Z/p,\mathbb{F}_p}F(\mathbb R^d,p)=\langle ab^{\frac{(d-1)(p-1)}{2}}, b^{\frac{(d-1)(p-1)}{2}+1}\rangle\subseteq \mathbb{F}_p[a,b]/\langle a^2\rangle.\]
# Existence of \(r\)-equilateral \(p\)-gons in \(\mathbb R^d\)
Assume that \(p\) is an odd prime number, \(r\geq 1\), \(d\geq 1\), \(\rho_1, \rho_2,\ldots,\rho_r\) are \(r\) metrics on \(\mathbb{R}^d\), each \(\rho_i\) equivalent to the Euclidean metric of \(\mathbb{R}^d\), and consider a sequence \(\{t_1,t_2,\ldots, t_r\}\), where \(t_i\in\{1,2,\ldots,(p-1)/2\}\). The problem is to show that there are \(p\) distinct points \(x_0,x_1\ldots,x_{p-1}\) in \(\mathbb{R}^d\) such that \(C_{t_i}\) is \(\rho_i\)-regular for all \(i=1,2,\ldots,r\). The configuration space for our problem is the configuration space of \(p\) distinct points in \(\mathbb{R}^d\), \(F(\mathbb{R}^d,p)=\{(x_0,x_1,\ldots, x_{p-1})\in(\mathbb{R}^d)^p:x_i\neq x_j, i\neq j\}\). Consider \(\mathbb{Z}/p=\langle\omega|\ \omega^p=1\rangle\) acting on \((\mathbb{R}^d)^p\) by \[\omega\cdot (x_0, x_1,\ldots,x_{p-1})=(x_{1}, x_{2},\ldots,x_{p-1},x_{0}).\] The group \(\mathbb{Z}/p\) acts freely on \(F(\mathbb{R}^d,p)\). Next, we define \(V=\mathbb{R}^p\times \cdots \times \mathbb{R}^p\), the product of \(r\) copies of \(\mathbb{R}^d\), and define our test map \(f=(f_1,f_2,\ldots,f_{r}):F(\mathbb{R}^d,p)\to V\), where \[f_i(x_0,\ldots,x_{p-1})=(\rho_i(x_0,x_{t_i}), \rho_i(x_{t_i},x_{2t_i}), \ldots, \rho_i(x_{(p-1)t_i},x_0)).\] The group \(\mathbb{Z}/p\) acts on \(\mathbb{R}^p\) by \(\omega\cdot(y_0, y_1,\ldots,y_{p-1})=(y_{1}, y_{2},\ldots,y_{p-1},y_{0})\) and \(\mathbb{Z}/p\) acts on \(V\) diagonally. The map \(f:F(\mathbb{R}^d,p)\to V\) becomes \(\mathbb{Z}/p\)-equivariant. The existence of \(p\) distinct points \(x_0,x_1\ldots,x_{p-1}\) in \(\mathbb{R}^d\) such that \(C_{t_i}\) is \(\rho_i\)-regular for any \(i\) is equivalent to the existence of an element \((x_0,\ldots,x_{p-1})\in F(\mathbb{R}^d,p)\) such that for each \(i=1,\ldots,r\), \(f_i(x_0,\ldots,x_{p-1})\in \Delta\), where \(\Delta=\{(t,t,\ldots,t)\in \mathbb{R}^d:t\in\mathbb{R}\}\); this is equivalent to say that \(im\ f\cap \Delta^{r}\neq \varnothing\). If we assume that \(im\ f\cap\Delta^r=\varnothing\), then we get a \(\mathbb{Z}/p\)-equivariant map \(f:F(\mathbb{R}^d,p)\to V\setminus \Delta^r\). Note that the action of \(\mathbb{Z}/p\) on \(\mathbb{R}^p\setminus \Delta\) is free and so is the action of \(\mathbb{Z}/p\) on \(V\setminus \Delta^r\). Thus, we have the following:
Lemma [\[lemmaequilateral1\]](#lemmaequilateral1){reference-type="ref" reference="lemmaequilateral1"}, combined with Lemma [\[lemmaequilateral\]](#lemmaequilateral){reference-type="ref" reference="lemmaequilateral"}, completes the proof of Theorem [\[equilateral\]](#equilateral){reference-type="ref" reference="equilateral"}.
Now we list some consequences of Theorem [\[equilateral\]](#equilateral){reference-type="ref" reference="equilateral"}.
# Existence of equilateral \(p\)-gons on deformed spheres
Theorem [\[equilateral\]](#equilateral){reference-type="ref" reference="equilateral"} shows the existence of equilateral \(p\)-gons in \((\mathbb R^d,\rho)\) for \(p\) prime if \(p<d\); it does not apply to \(m\)-gons when \(m\) is not prime, in particular it does not apply to quadrilaterals. For proving the existence of quadrilaterals, Fadell-Husseini index with integer coefficients is needed. In fact, we have the following result of Blagojević and Ziegler (see ):
From this theorem, the existence of equilateral quadrilaterals follows for any metric \(\rho\) in \(\mathbb R^3\) equivalent to the Euclidean metric (take \(f\) as the inclusion \(S^2\hookrightarrow \mathbb R^3\)). The proof of this theorem relies on the use of Fadell-Husseini index with integer coefficients (It is shown in that the use of mod. 2 coefficients is not enough). We generalize this result of Blagojević and Ziegler in Theorem [\[pgonsdeformed\]](#pgonsdeformed){reference-type="ref" reference="pgonsdeformed"}. The problem is to determine when a pair \((d,p)\) is admissible. Consider a continuous and injective map \(f:S^{d}\to\mathbb R^{d+1}\) and metrics \(\rho_1,\rho_2,\ldots,\rho_{(p-1)/2}\) equivalent to the Euclidean metric on \(\mathbb R^{d+1}\). Consider the map \(F=(f_1,f_2,\ldots,f_r):(S^{d})^p\to (\mathbb{R}^p)^{(p-1)/2}\) where \(f_j(x_0,x_1,\ldots,x_{p-1})\) equals \[(\rho_j(f(x_0), f(x_{1})), \rho_j(f(x_{j}), f(x_{2j})),\ldots,\rho_j(f(x_{(p-1)j}), f(x_0)),\] for \((x_0,x_1,\ldots,x_{p-1})\in (S^d)^{p}\), \(j=1,2,\ldots,(p-1)/2\). Let \(L=\{(x,\ldots,x)\in (S^{d})^p:x\in S^{d}\}\) and \(\Delta=\{(t,\ldots,t)\in \mathbb{R}^p: t\in\mathbb{R}\}\). Let us restrict \(F\) to \(X_{d,p}=(S^{d})^p\setminus L\). We assume that \(im\ F\cap \Delta^{(p-1)/2}=\varnothing\) so that we can consider \(F\) as a map \(F:X_{d,p}\to Y_{p}:=(\mathbb{R}^p)^{(p-1)/2}\setminus\Delta^{(p-1)/2}\). The group \(\mathbb{Z}/p\) acts on these spaces by permuting coordinates and making the map \(F\) equivariant. The orthogonal projection \(Y_p\to \Delta^{(p-1)/2}\setminus \{0\}\) is \(\mathbb Z/p\)-equivariant, and so is the normalization map \(\Delta^{(p-1)/2}\setminus \{0\}\to S(\Delta^{(p-1)/2})^{\perp}\), where \(S(\Delta^{(p-1)/2})^{\perp}\) is the unit sphere of \((\Delta^{(p-1)/2})^{\perp}\). By composing, we get an equivariant map \(Y_{p}\to S(\Delta^{(p-1)/2})^{\perp}\cong S^{(p-1)^2/1-1}\).
We are going to consider the following problem, which is more general: to find conditions on a triplet \((d,r,p)\) such that there is no \(\mathbb{Z}/p\)-equivariant map \(X_{d,p}\to S(\Delta^r)^{\perp}\).
## The index of the sphere \(S(\Delta^{r})^{\perp}\)
The unit sphere \(S(\Delta^{r})^{\perp}\) is a sphere of dimension \(r(p-1)-1\) with a free action of \(\mathbb{Z}/p\). Then, by Proposition [\[indexzivaljevic\]](#indexzivaljevic){reference-type="ref" reference="indexzivaljevic"}, \(ii)\), we have \[Ind_{\mathbb{Z}/p,\mathbb{F}_p}S(\Delta^{r})^{\perp}=\langle b^{r(p-1)/2}\rangle\subseteq \mathbb{F}_p[a,b]/\langle a^2\rangle=H^{\ast}(B\mathbb{Z}/p;\mathbb{F}_p).\] We are not going to compute explicitly the index of \(X_{d,p}\), but rather we are going to give conditions on \(d,r,p\) such that the element \(b^{r(p-1)/2}\in H^{\ast}(B\mathbb{Z}/p;\mathbb{F}_p)\) does not belong to the index of \(X_{d,p}\). We do this by using the Serre spectral sequence of the fibration \(X_{d,p}\to (X_{d,p})_{\mathbb{Z}/p}\to B\mathbb{Z}/p\) to show that the element \(b^{r(p-1)/2}\in H^{\ast}(B\mathbb{Z}/p;\mathbb{F}_p)=E_2^{\ast,0}\) survives forever. This means that \(b^{r(p-1)/2}\) is not in \(Ind_{\mathbb{Z}/p,\mathbb{F}_p}X_{d,p}\). The \(E_2\)-page of the Serre spectral sequence of the fibration \(X_{d,p}\to (X_{d,p})_{\mathbb{Z}/p}\to B\mathbb{Z}/p\), \(E_2^{s,t}=H^s(B\mathbb{Z}/p; H^t(X_{d,p};\mathbb{F}_p))\) has to be understood as the cohomology of the group \(\mathbb{Z}/p\) with coefficients in the \(\mathbb{Z}/p\)-module \(H^{\ast}(X_{d,p};\mathbb{F}_p)\).
## The cohomology \(H^{\ast}(X_{d,p};\mathbb{F}_p)\) as a \(\mathbb{Z}/p\)-module
We will use Lefschetz duality (see, theorem 70.2, page 415):
We apply theorem [\[lefschetz\]](#lefschetz){reference-type="ref" reference="lefschetz"} to \(((S^d)^p, L)\) and \(G=\mathbb{Z}/p\) to have \[H^{\ast}(X_{d,p};\mathbb{F}_p)=H^{\ast}((S^{d})^p\setminus L;\mathbb{F}_p)=H_{pd-\ast}((S^{d})^p, L;\mathbb{F}_p).\] To compute the homology \(H_{\ast}((S^{d})^p, L;\mathbb{F}_p)\) we are going to use the long exact sequence of the pair \(((S^d)^p, L)\). Since \(L\cong S^{d}\), \(H_i(L;\mathbb{F}_p)=\mathbb{F}_p\) if \(i=0,d\) and \(H_i(L;\mathbb{F}_p)=0\) otherwise. An induction argument using the the Künneth isomorphism yields that the only nonzero homology groups \(H_i((S^{d})^p;\mathbb{F}_p)\) occurs when \(i=0,d,2d,\ldots,pd\); besides, \(H_{jd}((S^{d})^p;\mathbb{F}_p)\) is the direct sum of all terms of the form \(A_1\otimes A_2\otimes \cdots \otimes A_p\), where \(j\) of the \(A_l\)'s are equal to \(H_{d}(S^{d};\mathbb{Z}/p)\) and the remaining \(A_l\)'s are equal to \(H_0(S^{d};\mathbb{Z}/p)\). If \(N:=(\mathbb{F}_p)^{\oplus(p)}\) with the action of \(\mathbb{Z}/p\) given by permuting the \(p\) copies of \(\mathbb{F}_p\) in the obvious way, then we have \[H_{i}((S^{d})^p;\mathbb{F}_p)\cong \begin{cases} \mathbb{F}_p, & i=0,pd,\\ N^{\oplus \frac{1}{p}{p\choose i/d}}, & i=d,2d,\ldots, (p-1)d,\\ 0,& \text{otherwise.} \end{cases}\] The long exact sequence in homology of the pair \(((S^{d})^p, L)\) gives isomorphisms \[H_{jd}((S^{d})^p,L;\mathbb{F}_p)\cong H_{jd}((S^{d})^p;\mathbb{F}_p)\] for \(j=2,3,\ldots,p\), and an exact sequence \[0\rightarrow H_{d+1}((S^{d})^p,L)\rightarrow H_{d}(L)\overset{\iota_{\ast}}{\rightarrow}H_{d}((S^{d})^p)\rightarrow H_{d}((S^{d})^p,L)\rightarrow 0,\] where \(\iota_{\ast}\) is the map induced by the inclusion \(L\hookrightarrow (S^{d})^p\). We have that \(H_{i}((S^{d})^p,L;\mathbb{F}_p)=0\) in any other case. If \(x\) is the generator of \(H_{d}(L;\mathbb{F}_p)\) and \(x_i\) the generator of \(H_{d}((S^{d})^p;\mathbb{F}_p)\) carried by the \(i\)-th copy of \(S^{d}\) in \((S^{d})^p\), then we have that \(\iota_{\ast}(x)=x_1+\cdots+x_p\). Thus, \(\iota_{\ast}\) is injective, \(H_{d+1}((S^{d})^p,L;\mathbb{F}_p)=0\) and \[\begin{aligned}
H_{d}((S^{d})^p,L;\mathbb{F}_p)& \cong H_{d}((S^{d})^p;\mathbb{F}_p)/im\ \iota_{\ast}\\ &=\langle x_1,\ldots,x_p\rangle/\langle x_1+\cdots+x_p\rangle\\ &=:M.\\ \end{aligned}\] Thus, we have that \[H_i((S^d)^p,L;\mathbb{F}_p)= \begin{cases} M,& i=d,\\ N^{\oplus \frac{1}{p}{p\choose i/d}},& i=2d,\ldots, (p-1)d,\\ \mathbb{Z}/p,& i=pd,\\ 0,& \text{otherwise.} \end{cases}\] and by theorem [\[lefschetz\]](#lefschetz){reference-type="ref" reference="lefschetz"} and using that \({p\choose i/d}={p\choose p-i/d}\)): \[H^i(X_{d,p};\mathbb{F}_p)= \begin{cases} \mathbb{Z}/p,& i=0,\\ N^{\oplus \frac{1}{p}{p\choose i/d}},& i= d,\ldots, (p-2)d,\\ M,& t=(p-1)d,\\ 0,& \text{otherwise.} \end{cases}\]
## The Serre spectral sequence of the fibration \(X_{d,p}\to (X_{d,p})_{\mathbb{Z}/p}\to B\mathbb{Z}/p\)
The \(E_2\)-page of the Serre spectral sequence of the fibration \(X_{d,p}\to (X_{d,p})_{\mathbb{Z}/p}\to B\mathbb{Z}/p\) is given by: \[E_2^{s,t}=H^s(\mathbb{Z}/p;H^t(X_{d,p};\mathbb{F}_p)).\] We have that \(H^t(X_{d,p};\mathbb{F}_p)\neq 0\) only for \(t=0,d,2d,\ldots,(p-1)d\). For \(t=0\), \(H^0(X_{d,p};\mathbb{F}_p)=\mathbb{F}_p\) so that \(E_2^{\ast,0}=H^{\ast}(\mathbb{Z}/p;\mathbb{F}_p);\mathbb{F}_p[a,b]/\langle a^2\rangle\). To compute the cohomology groups \(H^{s}(\mathbb{Z}/p;N)\) we consider the *norm map* \(\nu: N\to N\) given by multiplication by \(1+\omega+\omega^2+\cdots +\omega^{p-1}\) (where \(\mathbb Z/p=\langle \omega|\omega^p=1\rangle\)) and have that for \(s\geq 2\) even \(H^{s}(\mathbb{Z}/p;N)=coker\ \overline{\nu}\) and for \(s\) odd \(H^{s}(\mathbb{Z}/p;N)=ker\ \overline \nu\), where \(\overline \nu: N_{\mathbb Z/p}\to N^{\mathbb Z/P}\) is induced by \(\nu\) (see ).This map \(\overline \nu\) is an isomorphism, so we get the following: \[H^{s}(\mathbb{Z}/p;N)= \begin{cases} \mathbb{Z}/p, & s=0,\\ 0,& s>0. \end{cases}\] For \(t=d,2d,\ldots,(p-2)d\) we have \[E_2^{s,t}=H^s(\mathbb{Z}/p;N^{\oplus\frac{1}{p}{p\choose t/p}})= \begin{cases} (\mathbb{Z}/p)^{\frac{1}{p}{p\choose t/p}},& s=0,\\ 0,& s>0. \end{cases}\] We also have \(E_2^{s,(p-1)d}=H^s(\mathbb{Z}/p;M)\). Otherwise, \(E_2^{s,t}=0\). The \(E_2\)-term of the spectral sequence is shown in figure [\[spectralsequence1\]](#spectralsequence1){reference-type="ref" reference="spectralsequence1"}. The element \(b^{r(p-1)/2}\) lives in \(E_2^{r(p-1),0}=H^{r(p-1)}(\mathbb{Z}/p;\mathbb{F}_p)\). With our description of \(E_2^{\ast,\ast}\) we see that there are only two differentials that might hit elements in \(E_2^{r(p-1),0}\), that are \(d_{r(p-1)}:E_2^{0, r(p-1)-1}\to E_2^{r(p-1),0}\) and \(d_{(p-1)d+1}:E_2^{(r-d)(p-1)-1, d(p-1)}\to E_2^{r(p-1),0}\).
If \(r\leq d\), then \(E_2^{(2r-d)(p-1)-1, d(p-1)}=0\). On the other hand, \(E_2^{0, r(p-1)-1}\neq 0\) implies that \(r(p-1)-1=jd\) for some \(j=1,2,\ldots, p-1\). The result is that if \(r\leq d\) and \(r(p-1)\) is not of the form \(jd+1\) for any \(j=1,2,\ldots,p-1\), then the two differentials above are zero. This implies that \(b^{r(p-1)/2}\) is not in the index \(Ind_{\mathbb{Z}/p,\mathbb{Z}/p}X_{d,p}\). Thus we have:
Theorem [\[pgonsdeformed\]](#pgonsdeformed){reference-type="ref" reference="pgonsdeformed"} follows from proposition [\[admissibleequivariant\]](#admissibleequivariant){reference-type="ref" reference="admissibleequivariant"} and proposition [\[triplets\]](#triplets){reference-type="ref" reference="triplets"} by taking \(r=(p-1)/2\). A direct consequence of Theorem [\[pgonsdeformed\]](#pgonsdeformed){reference-type="ref" reference="pgonsdeformed"} is: | {'timestamp': '2017-06-07T02:03:32', 'yymm': '1706', 'arxiv_id': '1706.01618', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01618'} |
null | null |
# Introduction
Our motivation to work with massive spin-2 particles in a curved background is twofold. On one hand, they can represent massive gravitons at the linearized approximation, on the other hand, they can be understood as elementary massive spin-2 particles in a given gravitational background. Regarding the motivation for massive gravitons, they lead to an weaker gravitational interaction at large distances, which could contribute to the observed accelerated expansion of the universe at large distances. Although the recent detection of gravitational waves is consistent with massless gravitons, predicted by the usual (massless) general relativity, massive gravitons are not ruled out. The mentioned experiment sets an upper bound of about \(10^{-22} \, eV\) for the graviton mass. Furthermore, previous theoretical obstacles for massive gravitons like the vDVZ mass discontinuity and the existence of ghosts in the nonlinear theory have been tackled by the addition of fine tuned nonlinear self-interaction terms for the metric fluctuation, see and the bimetric model of. Those models are based on previous ideas of and and have recently led to intense work on massive gravity and related topics, see the review works. Regarding elementary massive spin-2 particles, the coupling of higher spin particles to electromagnetic and gravitational interactions is a longstanding problem. Since any elementary particle must couple to gravity, one first needs to check the gravitational interaction as in and. Usually, unitarity and causality are lost in interacting theories of higher spin particles. Those particles require the use of higher rank tensors which have too many components. The redundant components must vanish on shell. They work like auxiliary fields. However, when interactions are turned on, some of those auxiliary fields may become dynamic, giving rise to negative contributions to the Hamiltonian (instabilities) and incorrect number of degrees of freedom. Basically all studies of interacting massive spin-2 particles and the modern massive gravity theories, as, start with the paradigmatic free theory suggested by Fierz and Pauli in. It describes massive spin-2 particles via a symmetric and traceful rank-2 tensor \(h_{\mu\nu} = h_{\nu\mu}\). It is the metric fluctuation in massive gravitational theories, \(g_{\mu\nu} = \eta_{\mu\nu} + h_{\mu\nu}\). A natural question concerns the independence of the outcome of such studies on the underlying specific massive spin-2 model. In we have started with a rather general second order (in derivatives) Ansatz for a quadratic Lagrangian for a nonsymmetric rank-2 tensor \(e_{\mu\nu}\) and by requiring the existence of only one massive physical pole in the spin-2 sector of the propagator we have obtained three families of consistent free theories describing massive spin-2 particles. One of them is the usual Fierz-Pauli (FP) family which includes the FP model written in terms of a symmetric tensor. The other two families require a nonsymmetric tensor. There is no local field redefinition relating those families in general. One of the families is given in ([\[La1\]](#La1){reference-type="ref" reference="La1"}) and the other one in ([\[Lc\]](#Lc){reference-type="ref" reference="Lc"}). They depend on an arbitrary real constant, \(a_1\) and \(c\), respectively. See also for the special case \(a_1 =-1/4\). Here we couple a background gravitational field to those theories by including also nonminimal terms and look for curved space generalization of the tensor, vector and scalar constraints which are necessary for getting rid of nonphysical degrees of freedom. We require that the coefficients of the nonmininal terms be analytic functions of \(m^2\). Such restriction plays a key role in our work and leads us to constraint the gravitational background to Einstein spaces, see further comments in the conclusion. In section [2](#sectionLa1){reference-type="ref" reference="sectionLa1"} we deal with \({\cal L}(a_1)\), while in section [3](#sectionlc){reference-type="ref" reference="sectionlc"} we study the \({\cal L}_{nFP}(c)\) case. In section [4](#concl){reference-type="ref" reference="concl"} we draw our conclusions. In the appendix we briefly show the technical difficulties in arbitrary backgrounds.
# Family of Lagrangians \(\mathcal{L}(a_1)\) {#sectionLa1}
## Main results in the flat space
In the family of second order Lagrangians \(\mathcal{L}(a_1)\) has been presented in arbitrary dimensions \(D\geq 3\), but here we focus in \(D=4\). It describes massive "spin-2" particles via a nonsymmetric rank-2 tensor \(e_{\mu\nu}\neq e_{\nu\mu}\) in the flat space[^1] for any value of the constant \(a_1\): \[\begin{aligned}
\mathcal{L}(a_1)=-\frac{1}{2}\partial^\mu e^{(\alpha\beta)}\partial_\mu e_{(\alpha\beta)}+\biggl(a_1+\frac{1}{4}\biggr)\partial^\mu e[\partial_\mu e-2\partial^\alpha e_{(\alpha\mu)}]+\nonumber \\ +[\partial^\alpha e_{(\alpha\beta)}]^2+\biggl(a_1-\frac{1}{4}\biggr)(\partial^\alpha e_{\alpha\beta})^2-\frac{m^2}{2}(e_{\mu\nu}e^{\nu\mu}-e^2) \label{La1} \end{aligned}\] We recover the FP theory at \(a_1=1/4\) where \(e_{[\mu\nu]}\) becomes non dynamic and it can be neglected. However there is no local field redefinition which takes us from the FP theory to \(a_1\neq 1/4\). The massless theory \(\mathcal{L}_{m=0}(a_1)\) is unitary in the ranges \(a_1\geq 1/4\) and \(a_1\leq-1/12\), it describes massless spin-2 particles plus a scalar field, except at \(a_1=1/4\) and \(a_1=-1/12\) where the scalar field disappears. At \(a_1=-1/12\) the model \(\mathcal{L}(a_1)\) intersects the nFP (non Fierz-Pauli) family of section [3](#sectionlc){reference-type="ref" reference="sectionlc"} at \(c=-1\), see ([\[Lc\]](#Lc){reference-type="ref" reference="Lc"}). The flat space equations of motion \(E_{\mu\nu}\equiv \dfrac{\delta S(a_1)}{\delta e^{\mu\nu}}=0\) are given by: \[\begin{aligned}
E_{\mu\nu}&=&\square e_{(\mu\nu)}+2\biggl(a_1+\frac{1}{4}\biggr)[\eta_{\mu\nu}(\partial^\alpha \partial^\beta e_{\alpha\beta}-\square e)+\partial_\mu\partial_\nu e]-\partial_\mu\partial^\alpha e_{(\alpha\nu)}+\nonumber \\ &\, &-\partial_\nu\partial^\alpha e_{(\alpha\mu)}-2\biggl(a_1-\frac{1}{4}\biggr)\partial_\mu \partial^\alpha e_{\alpha\nu}+m^2(\eta_{\mu\nu}e-e_{\nu\mu})=0 \end{aligned}\] From \(\partial^\nu E_{\mu\nu}=0\), we have the vector constraint: \[\begin{aligned}
\partial^\alpha e_{\alpha\nu}=\partial_\nu e \label{la1eq01} \end{aligned}\] Plugging ([\[la1eq01\]](#la1eq01){reference-type="ref" reference="la1eq01"}) back in \(E_{\mu\nu}\) we have, from \(E_{\mu\nu}-E_{\nu\mu}=0\), the tensor constraint: \[\begin{aligned}
e_{[\mu\nu]}=0 \label{fpc1} \end{aligned}\] From ([\[la1eq01\]](#la1eq01){reference-type="ref" reference="la1eq01"}) and ([\[fpc1\]](#fpc1){reference-type="ref" reference="fpc1"}) back in \(\eta^{\mu\nu}E_{\mu\nu}=0\), we obtain the final scalar constraint: \[\begin{aligned}
e=0 \label{fpc2} \end{aligned}\] and, consequently, from ([\[la1eq01\]](#la1eq01){reference-type="ref" reference="la1eq01"}) we have the transverse condition: \[\begin{aligned}
\partial^\alpha e_{\alpha\nu}=0 \label{fpc3} \end{aligned}\] The equations of motion \(E_{\mu\nu}=0\) become the Klein-Gordon equations \[\begin{aligned}
(\square-m^2)e_{(\mu\nu)}=0 \end{aligned}\] The FP conditions ([\[fpc1\]](#fpc1){reference-type="ref" reference="fpc1"}), ([\[fpc2\]](#fpc2){reference-type="ref" reference="fpc2"}) and ([\[fpc3\]](#fpc3){reference-type="ref" reference="fpc3"}) guarantee the correct number of 5 degrees of freedom consistent with \(5=2s+1\), see for a recent derivation of the FP conditions from first principles.
## Generalization of \(\mathcal{L}(a_1)\) to curved spaces
### General setup and constraints
If we want to construct a theory of massive "spin-2" field in a curved space out of a nonsymmetric rank-2 tensor we should provide the same number of propagating degrees of freedom as in the flat case. They correspond to the curved space version of the 11 Fierz-Pauli conditions ([\[fpc1\]](#fpc1){reference-type="ref" reference="fpc1"}), ([\[fpc2\]](#fpc2){reference-type="ref" reference="fpc2"}), ([\[fpc3\]](#fpc3){reference-type="ref" reference="fpc3"}), namely \(e_{[\mu\nu]}=0\), \(g^{\mu\nu}e_{\mu\nu}=0\) and \(\nabla^\mu e_{\mu\nu}=0\). Thus, from the 16 components of \(e_{\mu\nu}\), we end up with \(16-11=5\) degrees of freedom. Our calculations focus on the \(D=4\) case, but it can be generalized to \(D\) dimensions (\(D\geq 3\)). Generalizing ([\[La1\]](#La1){reference-type="ref" reference="La1"}) to curved spacetime we substitute all derivatives by the covariant ones and add nonminimal terms containing the curvature tensor as in the FP case. They also take care of ordering ambiguities. Requiring a quadratic theory in derivatives, consistent with the flat limit ([\[La1\]](#La1){reference-type="ref" reference="La1"}) and at most linear in curvatures, the most general action has the form[^2] \[\begin{aligned}
\mathcal{L}^g(a_1) &=&-\frac{1}{4}\nabla^\mu e^{\alpha\beta}\,\nabla_\mu e_{\alpha\beta}-\frac{1}{4}\nabla^\mu e^{\alpha\beta}\,\nabla_\mu e_{\beta\alpha}+a_1\nabla^\alpha e_{\alpha\beta}\,\nabla_\mu e^{\mu\beta}+\frac{1}{2}\nabla^\alpha e_{\alpha\beta}\,\nabla_{\mu}e^{\beta\mu}+\nonumber \\ &\,&+\frac{1}{4}\nabla^\alpha e_{\beta\alpha}\,\nabla_{\mu}e^{\beta\mu} +\biggl(a_1+\frac{1}{4}\biggr)\nabla^\mu e\,\nabla_\mu e-\biggl(a_1+\frac{1}{4}\biggr)\nabla^\mu e\,\nabla^\alpha \, e_{\alpha\mu}+\nonumber \\ &\,&-\biggl(a_1+\frac{1}{4}\biggr)\nabla^\mu e \,\nabla^{\alpha}e_{\mu\alpha}-\frac{m^2}{2}(e_{\alpha\beta}\,e^{\beta\alpha}-e^2)+f_1\,R\,e^{\alpha\beta}\,e_{\alpha\beta}+f_2\,R\,e^2+\nonumber \\ &\,& +f_3\,R_{\alpha\beta\mu\nu}\,e^{\alpha\mu}\,e^{\beta\nu}+f_4\,R_{\alpha\beta}\,e^{\alpha\mu}\,{e^{\beta}}_\mu +f_5\,R_{\alpha\beta}\,e^{\alpha\beta}\,e+f_6\,R_{\alpha\beta\mu\nu}\,e^{\alpha\beta}\,e^{\mu\nu}+\nonumber \\ &\,& +f_7\,R_{\alpha\beta}\,e^{\alpha\mu}\,{e_\mu}^\beta +f_8\,R\,e^{\alpha\beta}\,e_{\beta\alpha}+f_9\,R_{\alpha\beta}\,e^{\mu\alpha}\,{e_{\mu}}^\beta \label{La1g} \end{aligned}\] where \(f_j\) \((j=1,2...,9)\) are arbitrary constants for the time being. Varying the action with respect to \(e^{\rho\sigma}\), we obtain the equations of motion in curved space: \[\begin{aligned}
E_{\rho\sigma}\doteq\frac{\delta S}{\delta e^{\rho\sigma}} &=& \frac{1}{2}\square(e_{\rho\sigma}+e_{\sigma\rho})-2a_1\nabla_{\rho}\nabla^{\mu}e_{\mu\sigma}-\frac{1}{2}\nabla_{\rho}\nabla^\mu e_{\sigma\mu}-\frac{1}{2}\nabla_\sigma\nabla^\mu e_{\mu\rho}-\frac{1}{2}\nabla_\sigma \nabla^\mu e_{\rho\mu}+\nonumber \\ &\,&+2\biggl( a_1+\frac{1}{4}\biggr)\biggl[-g_{\rho\sigma}\square e +\, \nabla_\rho\nabla_\sigma e +g_{\rho \sigma}\frac{\nabla^{\mu}\nabla^\alpha(e_{\alpha\mu}+e_{\mu\alpha})}{2}\biggr]+ \nonumber \\ &\,&-m^2(e_{\sigma\rho}-e\,g_{\rho\sigma})+2f_1\, R\,e_{\rho\sigma}+2f_2\,R\,g_{\rho\sigma}\,e+2f_3\,R_{\rho\beta\sigma\nu}\,e^{\beta\nu}+2f_4\,R_{\rho\beta}\,{e^{\beta}}_\sigma +\nonumber \\ &\,& +f_5\, R_{\rho \sigma}\, e +f_5\,R_{\alpha\beta}\,g_{\rho\sigma}\,e^{\alpha\beta}+2f_6\,R_{\alpha\beta\rho\sigma}\,e^{\alpha\beta}+f_7\,R_{\alpha\sigma}\,{e^{\alpha}}_\rho+f_7\, R_{\rho\alpha}\,{e_\sigma}^\alpha +\nonumber \\ &\,& +2f_8\,R\,e_{\sigma\rho}+2f_9\,R_{\sigma\beta}\,{e_\rho}^\beta =0 \label{eom} \end{aligned}\] By applying one derivative on the equations of motion and after several manipulations, one obtains: \[\begin{aligned}
\mathcal{C}_\rho\doteq\nabla^\sigma E_{\rho\sigma} &=& +(1-2f_3-2f_6)R_{\rho\lambda\sigma\alpha}\nabla^\alpha e^{\lambda\sigma}+(1+2f_6)R_{\rho\lambda\sigma\alpha}\nabla^\alpha e^{\sigma\lambda}+\nonumber \\ &\,& +\biggl(\frac{1}{2}+f_7\biggr) R^{\lambda\alpha}\nabla_\alpha e_{\lambda\rho}+\biggl(\frac{1}{2}+2f_9\biggr)R^{\lambda\alpha}\nabla_\alpha e_{\rho\lambda}+(f_7-2a_1)R_{\lambda\rho}\nabla_\mu e^{\mu\lambda}+\nonumber \\ &\,& +\biggl(2f_4-\frac{1}{2}\biggr)R_{\lambda\rho}\nabla_\mu e^{\lambda\mu}+2f_1R\nabla^\sigma e_{\rho\sigma}+\biggl(\frac{1}{4}+2f_1+f_9\biggr)e_{\rho\sigma}\nabla^\sigma R+\nonumber \\ &\,& +\biggl(2f_2+\frac{f_5}{2}\biggr)e\nabla_\rho R+(m^2+2f_2R)\nabla_\rho e+\biggl(\frac{1}{2}+2f_4+2f_6\biggr)e^{\beta\sigma}\nabla_\sigma R_{\rho\beta}+\nonumber \\ &\,& +\biggl(\frac{1}{2}+2a_1+f_5\biggr)R_{\rho\sigma}\nabla^\sigma e-(1-2f_3-f_5)e^{\alpha\beta}\nabla_\rho R_{\alpha\beta}+f_5R_{\alpha\beta}\nabla_\rho e^{\alpha\beta}+\nonumber \\ &\,&-\biggl(\frac{1}{2}+2f_3+2f_6-f_7\biggr) e^{\sigma\alpha}\nabla_\sigma R_{\rho\alpha}+\biggl(\frac{1}{4}+\frac{f_7}{2}+2f_8\biggr)e_{\sigma\rho}\nabla^\sigma R+\nonumber \\ &\,&-(m^2-2f_8R)\nabla^\sigma e_{\sigma\rho} \label{vv} \end{aligned}\] Now, we define the tensor \(\mathcal{C}_{\rho\sigma}\): \[\begin{aligned}
\mathcal{C_{\rho\sigma}} &\doteq & E_{\rho\sigma}-E_{\sigma \rho}=\biggl(-2a_1+\frac{1}{2}\biggr)(\nabla_\rho\nabla^\mu e_{\mu\sigma}-\nabla_\sigma \nabla^\mu e_{\mu\rho})+\nonumber \\ &\,& +[m^2+2R(f_1-f_8)](e_{\rho\sigma}-e_{\sigma\rho})+2f_3\,R_{\rho\beta\sigma\nu}(e^{\beta\nu}-e^{\nu\beta})+\nonumber \\ &\, &+2f_4(R_{\rho\beta}\, {e^\beta}_\sigma-R_{\sigma\beta}\, {e^\beta}_\rho )+4f_6\,R_{\alpha\beta\rho\sigma} \,e^{\alpha\beta}+f_7 {R^\alpha}_\sigma(e_{\alpha\rho}-e_{\rho\alpha})+\nonumber \\ &\,& +f_7 {R_\rho}^\alpha(e_{\sigma\alpha}-e_{\alpha\sigma})+2f_9(R_{\sigma\beta}\, {e_\rho}^\beta-R_{\rho\beta}\, {e_{\sigma}}^\beta) \label{vt} \end{aligned}\] In order to find a scalar constraint we have to consider the most general scalar combination of the equations of motion \[\begin{aligned}
\mathcal{C}\doteq (b_0\, R\, +b_1m^2)g^{\rho\sigma}E_{\rho\sigma}+b_2 R^{\rho\sigma}E_{\rho\sigma}+b_3\nabla^\rho\nabla^\sigma E_{\rho\sigma} \label{vecompleto} \end{aligned}\] where \(b_j\) \((j=0,1,2,3)\) are arbitrary constants for now. By manipulating and simplifying as much as possible, we obtain the following expression: \[\begin{aligned}
\mathcal{C} = &\,&+[b_3(2f_4+f_7)-b_2]R^{\lambda\alpha}\nabla_\alpha\nabla^\rho e_{\lambda \rho}+2b_3(1-f_3)R_{\rho\lambda\sigma\alpha}\,\nabla^\rho\nabla^\alpha e^{\lambda\sigma}+\nonumber \\ &\,&-\Biggl[2b_2\biggl(\frac{1}{4}+a_1\biggr)-b_3\biggl(\frac{1}{2}-2a_1+f_7+2f_9\biggr)\Biggr]R^{\lambda\alpha}\nabla_\alpha \nabla^\rho e_{\rho\lambda}+\nonumber \\ &\, & +\Biggl[(b_0R+b_1m^2)\biggl(\frac{1}{2}+6a_1\biggr)+2b_2\biggl(\frac{1}{4}+a_1\biggr)R-b_3(m^2-2f_1R-2f_8R)\Biggr]\nabla^\lambda\nabla^\rho e_{\rho\lambda}+\nonumber \\ &\,&-\Biggl[(b_0R+b_1m^2)\biggl(\frac{1}{2}+6a_1\biggr) +2b_2\biggl(\frac{1}{4}+a_1\biggr)R-b_3(m^2+2f_2R)\Biggr]\square e+\nonumber \\ &\,&+(b_2+b_3f_5)R_{\lambda\rho}\, \square e^{\lambda \rho}+\Biggl[ b_2+b_3f_5+2\biggl(a_1-\frac{1}{4}\biggl)(b_2+b_3) +b_3\Biggr] R_{\lambda\rho}\nabla^\lambda\nabla^\rho e + \mathcal{C}_1\nonumber \\ \label{ve} \end{aligned}\] where \(\mathcal{C}_1\) contains up to first derivatives of \(e_{\rho\sigma}\). The expression ([\[ve\]](#ve){reference-type="ref" reference="ve"}) has seven terms with second derivatives of \(e_{\rho\sigma}\), which must be eliminated in order to become a scalar constraint. In the special case of the FP theory (\(a_1=1/4\)), the last two terms with second derivatives can only be simultaneously canceled if \(b_2=b_3=0\). Back in the other terms we need \(b_0 R + m^2 \, b_1=0\). However in this case we have no constraint whatsoever. This is in agreement with where the authors have chosen Einstein spaces in order to surmount such difficulty. In the general case \(a_1 \ne 1/4\) we have to find a solution for the system below: \[\begin{aligned}
b_3(2f_4+f_7)-b_2=0\\ b_3(1-f_3)=0\\ 2b_2\biggl(\frac{1}{4}+a_1\biggr)-b_3\biggl(\frac{1}{2}-2a_1+f_7+2f_9\biggr)=0 \\ (b_0R+b_1m^2)\biggl(\frac{1}{2}+6a_1\biggr)+2b_2\biggl(\frac{1}{4}+a_1\biggr)R-b_3(m^2-2f_1R-2f_8R)=0 \label{systemLa1-1Ap}\\ (b_0R+b_1m^2)\biggl(\frac{1}{2}+6a_1\biggr) +2b_2\biggl(\frac{1}{4}+a_1\biggr)R-b_3(m^2+2f_2R)=0 \label{systemLa1-2Ap}\\ b_2+b_3f_5=0\\ 2b_2\biggl(\frac{1}{4}+a_1\biggl)+2b_3\biggl(\frac{1}{4}+a_1+\frac{f_5}{2}\biggl)=0 \label{systemLa1Ap} \end{aligned}\] Without restrictions in the background, as shown in the appendix, we have not been able to solve the previous system and get \(e=0\) from the scalar constraint. So, we are going to restrict the gravitational background to Einstein spaces[^3] as in the FP case, \[\begin{aligned}
R_{\mu\nu}=\frac{R}{4}g_{\mu\nu} \label{es} \end{aligned}\] Now we can rewrite ([\[vv\]](#vv){reference-type="ref" reference="vv"}), ([\[vt\]](#vt){reference-type="ref" reference="vt"}) and ([\[ve\]](#ve){reference-type="ref" reference="ve"}) as follows: \[\begin{aligned}
\mathcal{C}_\rho\doteq\nabla^\sigma E_{\rho\sigma} &=& (1-2f_3-2f_6)R_{\rho\lambda\sigma\alpha}\,\nabla^\alpha e^{\lambda\sigma}+(1+2f_6)\, R_{\rho\lambda\sigma\alpha}\nabla^\alpha e^{\sigma\lambda}+2\tilde{f_1 R\,\nabla^\lambda e_{\rho\lambda}+\nonumber \\ &\,&+\biggl[\frac{1}{2}\biggl(\frac{1}{4}-a_1+4\tilde{f_8}\biggl)R-m^2\biggr]\nabla^\lambda e_{\lambda\rho}+\biggl[\frac{1}{2}\biggl(\frac{1}{4}+a_1+4\tilde{f_2}\biggr)R+m^2\biggr]\nabla_\rho e=0\nonumber \\ \label{vvnovo} \end{aligned}\] \[\begin{aligned}
\mathcal{C_{\rho\sigma}}\doteq E_{\rho\sigma}-E_{\sigma \rho}&=&+ \biggl(-2a_1+\frac{1}{2}\biggr)(\nabla_\rho\nabla^\mu e_{\mu\sigma}-\nabla_\sigma \nabla^\mu e_{\mu\rho})+2(f_3+2f_6)\,R_{\rho\beta\sigma\nu}(e^{\beta\nu}-e^{\nu\beta})+\nonumber \\ &\,&+ \biggl[m^2+(2\tilde{f_1}-2\tilde{f_8})R\biggl](e_{\rho\sigma}-e_{\sigma\rho})=0 \label{vtnovo} \end{aligned}\] \[\begin{aligned}
\mathcal{C}& \doteq & \tilde{b}_1 g^{\rho\sigma}E_{\rho\sigma}+b_3\nabla^\rho\nabla^\sigma E_{\rho\sigma} \nonumber \\ &=& +2b_3(1-f_3)R_{\rho\lambda\sigma\alpha}\nabla^\rho\nabla^\alpha e^{\lambda\sigma}+ \nonumber \\ &\,&+\Biggl[-6\tilde{b}_1\biggl(a_1+\frac{1}{12}\biggr)+\frac{b_3}{2}\biggl(\frac{1}{4}+a_1+4\tilde{f_2}\biggr)R+b_3m^2\Biggr]\square e+\nonumber \\ &\, & +\Biggl[6\tilde{b}_1\biggl(a_1+\frac{1}{12}\biggr)+\frac{b_3}{2}\biggl(\frac{1}{4}-a_1+4\tilde{f_1}+4\tilde{f_8}\biggr)R-b_3m^2\Biggr]\nabla^\lambda\nabla^\rho e_{\rho\lambda}+\nonumber \\ &\,&+\tilde{b}_1\biggl[3m^2+\biggl(2\tilde{f_1}+8\tilde{f_2}+\frac{f_3}{2}+2\tilde{f_8}\biggr)R\biggr] e =0 \label{venovo} \end{aligned}\] Motivated by the substitution of ([\[es\]](#es){reference-type="ref" reference="es"}) back in ([\[La1g\]](#La1g){reference-type="ref" reference="La1g"}) and ([\[vecompleto\]](#vecompleto){reference-type="ref" reference="vecompleto"}) we have defined: \[\begin{aligned}
\tilde{f_1}&\doteq &f_1+\dfrac{f_4}{4}+\dfrac{f_9}{4}\nonumber \\ \tilde{f_2}&\doteq &f_2+\dfrac{f_5}{4}\label{ftilde} \\ \tilde{f_8}&\doteq &f_8+\dfrac{f_7}{4}\nonumber \\ \tilde{b}_1 &\doteq &b_0\, R + b_1 m^2+\dfrac{b_2R}{4}\nonumber \end{aligned}\] The expression ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}) is already a vector constraint since it does not have second derivatives of the field. It corresponds to four constraints, in total. The same does not occur in expressions ([\[vtnovo\]](#vtnovo){reference-type="ref" reference="vtnovo"}) and ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}). First, let us turn ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}) into a scalar constraint. We need to solve the system below: \[\begin{aligned}
b_3(1-f_3)=0\nonumber \\-6\tilde{b}_1\biggl(a_1+\frac{1}{12}\biggr)+\frac{b_3}{2}\biggl(\frac{1}{4}+a_1+4\tilde{f_2}\biggr)R+b_3m^2=0\label{sistema0}\\ 6\tilde{b}_1\biggl(a_1+\frac{1}{12}\biggr)+\frac{b_3}{2}\biggl(\frac{1}{4}-a_1+4\tilde{f_1}+4\tilde{f_8}\biggr)R-b_3m^2=0\nonumber \end{aligned}\] It is easy to see that the solution of ([\[sistema0\]](#sistema0){reference-type="ref" reference="sistema0"}) back in ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}) leads to the scalar constraint \(e=0\), provided the coefficient of \(e\) is different from zero in ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}). However, the expression ([\[vtnovo\]](#vtnovo){reference-type="ref" reference="vtnovo"}) still has terms with second derivatives. For these terms to be cancelled[^4], it's necessary that \(\nabla^\mu e_{\mu\nu}=0\). We can get this from the vector constraint ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}) if an appropriate choice of parameters is made. More specifically, since the solution of ([\[sistema0\]](#sistema0){reference-type="ref" reference="sistema0"}) requires \(f_3=1\), if we set \(\tilde{f_1}=0\) and \(f_6=-\frac{1}{2}\), we obtain automatically from ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}) that \(\nabla^\mu e_{\mu\nu}=0\) as far as the coefficient of \(\nabla^\mu e_{\mu\nu}\) in ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}) is non-null. The solution of the system given in ([\[sistema0\]](#sistema0){reference-type="ref" reference="sistema0"}) with the additional equations \(\tilde{f_1}=0\) and \(f_6=-\frac{1}{2}\) is given by: \[\begin{aligned}
f_3 &=& 1 \nonumber \\ \tilde{f_8} &=&-\frac{1}{8}-\tilde{f}_2\label{solution1}\\ \tilde{b}_1 &=& \frac{b_3}{1+12a_1}\Biggl[2m^2+\biggl(\dfrac{1}{4}+a_1+4\tilde{f_2}\biggr)R\Biggr]\nonumber \end{aligned}\] Returning this solution in ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}), ([\[vtnovo\]](#vtnovo){reference-type="ref" reference="vtnovo"}) and ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}) we finally get all necessary constraints. More specifically, from ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}) we obtain the scalar constraint: \[\begin{aligned}
e=0 \label{vea1} \end{aligned}\] Using ([\[solution1\]](#solution1){reference-type="ref" reference="solution1"}) and the result \(e=0\) in ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}), we have the vector constraint: \[\begin{aligned}
\nabla^\sigma e_{\sigma\rho}=0 \label{vva1} \end{aligned}\] Finally, using ([\[solution1\]](#solution1){reference-type="ref" reference="solution1"}) and the results ([\[vea1\]](#vea1){reference-type="ref" reference="vea1"}) and ([\[vva1\]](#vva1){reference-type="ref" reference="vva1"}) in ([\[vtnovo\]](#vtnovo){reference-type="ref" reference="vtnovo"}), we achieve the tensor constraint: \[\begin{aligned}
e_{[\rho\sigma]}=0 \label{vta1} \end{aligned}\] once its coefficient in ([\[vtnovo\]](#vtnovo){reference-type="ref" reference="vtnovo"}) is nonvanishing too. Summarizing, we have found all the FP constraints: \[\begin{aligned}
e=0\label{vinculosa1-1}\\ \nabla^\sigma e_{\sigma\rho}=0\label{vinculosa1-2} \\ e_{[\rho\sigma]}=0 \label{vinculosa1-3} \end{aligned}\] if the restriction below are respected \[\begin{aligned}
b_3\tilde{m}^2\Biggl[2\tilde{m}^2+\biggl(-\dfrac{1}{4}+a_1\biggr)R\Biggr]\Biggl[3\tilde{m}^2-\frac{R}{2}\Biggr]\neq 0 \label{restrictionLa1} \end{aligned}\] \[\begin{aligned}
\tilde{m}^2\equiv m^2+\biggl(\frac{1}{4}+2\tilde{f}_2\biggr)R \label{mtildela1} \end{aligned}\] while the equations of motion become \[\begin{aligned}
E_{\rho\sigma}=(\square-\tilde{m}^2)e_{\rho\sigma}+2R_{\rho\alpha\sigma\beta}\ e^{\alpha\beta} \end{aligned}\] Therefore, we end up with \(16-11=5\) degrees of freedom, which is the correct count for a massive spin-2 particle (\(5=2s+1\)). The final curved space theory still contains 2 free parameters: \(\tilde{f_2}\) and \(a_1\), with the restriction ([\[restrictionLa1\]](#restrictionLa1){reference-type="ref" reference="restrictionLa1"}) and \((a_1+1/12)(a_1-1/4)\neq 0\). For sake of comparison with we focus now on a special subcase of Einstein spaces, namely the maximally symmetric spaces: \[\begin{aligned}
R_{\alpha\beta\rho\sigma}=\frac{R}{12}(g_{\alpha\rho}g_{\beta\sigma}-g_{\alpha\sigma}g_{\beta\rho}) \label{Riems} \end{aligned}\] All the results from the previous section can be brought consistently to the maximally symmetric spaces ([\[Riems\]](#Riems){reference-type="ref" reference="Riems"}). Using ([\[ftilde\]](#ftilde){reference-type="ref" reference="ftilde"}) and ([\[solution1\]](#solution1){reference-type="ref" reference="solution1"}), the lagrangian ([\[La1g\]](#La1g){reference-type="ref" reference="La1g"}) in a maximally symmetric background becomes \[\begin{aligned}
{\mathcal{L}}^{\mbox{\tiny{(MSS)}}}(a_1) &=&-\frac{1}{4}\nabla^\mu e^{\alpha\beta}\,\nabla_\mu e_{\alpha\beta}-\frac{1}{4}\nabla^\mu e^{\alpha\beta}\,\nabla_\mu e_{\beta\alpha}+a_1\nabla^\alpha e_{\alpha\beta}\,\nabla_\mu e^{\mu\beta}+\frac{1}{2}\nabla^\alpha e_{\alpha\beta}\,\nabla_{\mu}e^{\beta\mu}+\nonumber \\ &\,& +\frac{1}{4}\nabla^\alpha e_{\beta\alpha}\,\nabla_{\mu}e^{\beta\mu} +\biggl(a_1+\frac{1}{4}\biggr)\nabla^\mu e\,\nabla_\mu e-\biggl(a_1+\frac{1}{4}\biggr)\nabla^\mu e\,(\nabla^\alpha e_{\alpha\mu}+\nabla^\alpha e_{\mu\alpha})+\nonumber \\ &\,&-\frac{m^2}{2}(e_{\alpha\beta}\,e^{\beta\alpha}-e^2)-\frac{1}{24}\,R\,e^{\alpha\beta}\,e_{\alpha\beta}+\biggl(\tilde{f_2}+\frac{1}{12}\biggr)\,R\,e^2+\nonumber \\ &\,&-\frac{1}{4}\biggl(\frac{11}{12}+{a_1}+4\tilde{f_2}\biggr)\,R\,e^{\alpha\beta}\,e_{\beta\alpha} \label{lmssa1} \end{aligned}\] On the other hand, it has been presented in a model for massive spin-2 particles also with a nonsymmetric tensor \(e_{\mu\nu}\neq e_{\nu\mu}\) minimally coupled to maximally symmetric background. The lagrangian is known as dual massive gravity and is given by \[\begin{aligned}
{\mathcal{L}}^{\mbox{\tiny{(dual)}}} =\frac{1}{2}\nabla_\rho e_{\nu\sigma}(-\nabla^\rho e^{\nu\sigma}-\nabla^\nu e^{\rho\sigma}+\nabla^\nu e^{\sigma\rho}-\nabla^\rho e^{\sigma\nu}+\nabla^\sigma e^{\rho\nu}+\nabla^\sigma e^{\nu\rho})-m^2(e_{\mu\nu}e^{\nu\mu}-e^2)\nonumber \\ \label{acaoW} \end{aligned}\] As already discussed in, the model presented in is recovered from \({\mathcal{L}}^{\mbox{\tiny{(MSS)}}}(a_1)\) in the flat space when \(a_1=-1/4\). However, it's important to notice that the assumption of \(a_1=-1/4\) does not require a maximally symmetric space as we have shown here. The relation between \({\mathcal{L}}^{\mbox{\tiny{(dual)}}}\) and \({\mathcal{L}}^{\mbox{\tiny{(MSS)}}}(a_1)\) is given by \[\begin{aligned}
{\mathcal{L}}^{\mbox{\tiny{(MSS)}}}(a_1=-1/4)=\frac{1}{2}{\mathcal{L}}^{\mbox{\tiny{(dual)}}}-\biggl[\frac{1}{24}+\tilde{f_2}\biggr]R(e^{\alpha\beta}e_{\beta\alpha}-e^2) \end{aligned}\] Thus, the model of is a subcase of \({\mathcal{L}}^{\mbox{\tiny{(MSS)}}}(a_1=-1/4)\) where: \[\begin{aligned}
\tilde{f_2}=-\frac{1}{24} \label{c4} \end{aligned}\] With the above value of \(\tilde{f_2}\), the restrictions ([\[restrictionLa1\]](#restrictionLa1){reference-type="ref" reference="restrictionLa1"}) lead to two forbidden values for the scalar curvature, namely, \(R \ne-6\, m^2\) and \(R \ne 12\, m^2\). The first value differs by a sign from the restriction obtained in while the second one has not been mentioned. It is important to emphasize however, that \(\tilde{f_2}\) is a free parameter in the \(\mathcal{L}^g(a_1)\) model, so the inequality ([\[restrictionLa1\]](#restrictionLa1){reference-type="ref" reference="restrictionLa1"}) restricts the possible values of \(\tilde{f_2}\), not of the curvature \(R\). This happens because our original Lagrangian is more general than ([\[acaoW\]](#acaoW){reference-type="ref" reference="acaoW"}). In MSS there are no forbidden values for the scalar curvature in the \(\mathcal{L}^g(a_1)\) model for any value of \(a_1\), including \(a_1=-1/4\).
### Local symmetries of \(\mathcal{L}^g(a_1)\)
In the previous sections, we have found all the constraints of \(\mathcal{L}^g(a_1)\) model. The form of ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}) and ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}) suggests that some local symmetries of \(\mathcal{L}^g(a_1)\) may exist even in the massive case. For example, if the expression obtained for the vector constraint \(\nabla^\sigma E_{\rho\sigma}\) given in ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}) becomes identically null, instead of a vector constraint we would have four identities \(\nabla^\sigma E_{\rho\sigma}\equiv 0\) and, consequently, the theory acquires the vector symmetry \[\begin{aligned}
\delta e_{\rho\sigma}=\nabla_\sigma A_\rho \label{simla1} \end{aligned}\] since, up to a surface term, \[\begin{aligned}
\delta_{\mbox{\tiny{A}}} S=\int d^4x \ \frac{\delta S}{\delta e^{\rho\sigma}}\ \delta e^{\rho\sigma}=\int d^4 x \ E_{\rho\sigma} \ \nabla^\sigma A^\rho =-\int d^4x \ (\nabla^\sigma E_{\rho\sigma} )A_\rho =0 \end{aligned}\] In order that \(\nabla^\sigma E_{\rho\sigma}=0\) holds identically, see ([\[vvnovo\]](#vvnovo){reference-type="ref" reference="vvnovo"}), we need the conditions: \[\begin{aligned}
f_3 &=& 1\nonumber \\ \tilde{f_1}&=&0\nonumber \\ \tilde{f_8}&=&-\frac{1}{8}-\tilde{f_2}\label{condsimla1}\\ f_6&=&-\frac{1}{2}\nonumber \\ R&=&-\frac{8m^2}{1+4a_1+16\tilde{f_2}}\nonumber \end{aligned}\] We have checked explicitly that ([\[simla1\]](#simla1){reference-type="ref" reference="simla1"}) is indeed a symmetry of \(\mathcal{L}^g(a_1)\) if we use \((\ref{condsimla1})\). Analogously, we could find scalar symmetries for \(\mathcal{L}^g(a_1)\). The action is invariant under the transformation \[\begin{aligned}
{\delta e_{\rho\sigma}}^{\mbox{\tiny{(1)}}}=\nabla_\rho\nabla_\sigma \lambda \end{aligned}\] where \(\lambda\) is an arbitrary scalar, provided the conditions below are satisfied: \[\begin{aligned}
f_3&=&1\nonumber \\ \tilde{f_8}&=&-\frac{1}{8}-\tilde{f_1}-\tilde{f_2}\label{cond_sim1} \\ R&=&-\frac{8m^2}{1+4a_1+16\tilde{f_2}}\nonumber \end{aligned}\] In addition, there is another possible scalar symmetry which is \[\begin{aligned}
{\delta e_{\rho\sigma}}^{\mbox{\tiny{(2)}}}=12\, \nabla_\rho\nabla_\sigma \lambda +R\, g_{\rho\sigma}\lambda \end{aligned}\] where \(\lambda\) is an arbitrary scalar and the conditions below are required: \[\begin{aligned}
f_3&=&1\nonumber \\ \tilde{f_8}&=&-\frac{1}{8}-\tilde{f_1}-\tilde{f_2}\label{cond_sim2}\\ R&=&-\frac{12m^2}{1+24\tilde{f_2}}\nonumber \end{aligned}\] We leave for a future work a detailed study of the special cases ([\[condsimla1\]](#condsimla1){reference-type="ref" reference="condsimla1"}), ([\[cond_sim1\]](#cond_sim1){reference-type="ref" reference="cond_sim1"}) and ([\[cond_sim2\]](#cond_sim2){reference-type="ref" reference="cond_sim2"}). The appearance of vector and scalar symmetries are usually connected with massless and partially massless theories, respectively.
# Family of Lagrangians \(\mathcal{L}_{\mbox{\tiny{nFP}}}(c)\) {#sectionlc}
## Main results in the flat space
Analogously to the \(\mathcal{L}(a_1)\) case, in we can find another family of second order Lagrangians \(\mathcal{L}_{\mbox{\tiny{nFP}}}(c)\) which describes massive "spin-2" particles via a nonsymmetric rank-2 tensor in \(D=4\) flat spaces[^5] \[\begin{aligned}
\mathcal{L}_{\mbox{\tiny{nFP}}}(c)=-\frac{1}{2}\partial^\mu e^{(\alpha\beta)}\partial_\mu e_{(\alpha\beta)}+\frac{1}{6}\partial_\mu e[\partial^\mu e-2\partial_\nu e^{(\nu\mu)}]+\nonumber \\ +[\partial^\alpha e_{(\alpha\beta)}]^2-\frac{1}{3}[\partial_\mu e^{\mu\nu}]^2-\frac{m^2}{2}(e_{\mu\nu}e^{\nu\mu}+c\, e^2) \label{Lc} \end{aligned}\] The real constant \(c\) is arbitrary and nFP stands for "non Fierz-Pauli" since we do not need to have \(c=-1\). In such special case, however, the model \(\mathcal{L}_{\mbox{\tiny{nFP}}}(c=-1)\) coincides with \(\mathcal{L}(a_1)\) at \(a_1=-\frac{1}{12}\). The massless case \(\mathcal{L}^{m=0}_{\mbox{\tiny{nFP}}}(c)\) has first appeared in and describes massless spin-2 particles. If \(c\neq-\frac{1}{4}\) the FP conditions can be derived from the equations of motion as follows: \[\begin{aligned}
E_{\mu\nu}&=&\square e_{(\mu\nu)}+\frac{\eta_{\mu\nu}}{3}(\partial^\alpha \partial^\beta e_{\alpha\beta}-\square e)+\frac{1}{3}\partial_\mu\partial_\nu e-\partial_\mu\partial^\alpha e_{(\nu\alpha)}+\nonumber \\ &\, &-\partial_\nu\partial^\alpha e_{(\alpha\mu)}+\frac{2}{3}\partial_\mu \partial^\alpha e_{\alpha\nu}-m^2(e_{\nu\mu}+c\, \eta_{\mu\nu}e)=0\label{eomlc} \end{aligned}\] From \(\partial^\nu E_{\mu\nu}=0\), we have \[\begin{aligned}
\partial^\nu e_{\nu\mu}+c\, \partial_\mu e=0 \label{lceq01} \end{aligned}\] Back in ([\[eomlc\]](#eomlc){reference-type="ref" reference="eomlc"}), we obtain from \(E_{\mu\nu}-E_{\nu\mu}=0\): \[\begin{aligned}
e_{[\mu\nu]}=0 \label{fpc1-1} \end{aligned}\] From \(\eta^{\mu\nu}E_{\mu\nu}=0\), we have: \[\begin{aligned}
m^2\biggl(c+\frac{1}{4}\biggr)e=0\Longrightarrow e=0 \label{fpc2-2} \end{aligned}\] and, consequently, from ([\[lceq01\]](#lceq01){reference-type="ref" reference="lceq01"}) now we have \[\begin{aligned}
\partial^\alpha e_{\alpha\nu}=0 \label{fpc3-3} \end{aligned}\] Thus, the equations of motion given in ([\[eomlc\]](#eomlc){reference-type="ref" reference="eomlc"}) become the Klein-Gordon equations: \[\begin{aligned}
(\square-m^2)e_{(\mu\nu)}=0 \label{eomlc2} \end{aligned}\] If \(c=-\frac{1}{4}\) the model \(\mathcal{L}_{\mbox{\tiny{nFP}}}(c)\) is invariant under Weyl transformations: \(\delta_{\mbox{\tiny{W}}}e_{\mu\nu}=\eta_{\mu\nu}\phi\). We can fix the gauge \(e=0\) and obtain all the FP conditions ([\[fpc1-1\]](#fpc1-1){reference-type="ref" reference="fpc1-1"}), ([\[fpc2-2\]](#fpc2-2){reference-type="ref" reference="fpc2-2"}) and ([\[fpc3-3\]](#fpc3-3){reference-type="ref" reference="fpc3-3"}) and the Klein-Gordon equations ([\[eomlc2\]](#eomlc2){reference-type="ref" reference="eomlc2"}).
## Generalization of \({\mathcal{L}}_{\mbox{\tiny{nFP}}}(c)\) to curved spaces
### General setup and constraints
For \({\mathcal{L}}_{\mbox{\tiny{nFP}}}(c)\) the procedure was analogous to that used for \(\mathcal{L}(a_1)\). The most general expression for \({\mathcal{L}}_{\mbox{\tiny{nFP}}}(c)\) is the following one: \[\begin{aligned}
{\mathcal{L}}_{\mbox{\tiny{nFP}}}^g(c)&=&-\frac{1}{4}\nabla^\mu e^{\alpha\beta}\nabla_\mu e_{\alpha\beta}-\frac{1}{4}\nabla^\mu e^{\alpha\beta}\nabla_\mu e_{\beta\alpha}-\frac{1}{12}\nabla^\alpha e_{\alpha\beta}\nabla_\lambda e^{\lambda\beta}+\frac{1}{2}\nabla^\alpha e_{\alpha\beta}\nabla_\lambda e^{\beta\lambda}+\nonumber \\ &\,&+\frac{1}{4}\nabla^\alpha e_{\beta\alpha} \nabla_\lambda e^{\beta\lambda}+\frac{1}{6}\nabla^\mu \nabla_\mu e-\frac{1}{3}\nabla^\alpha e_{\alpha\beta}\nabla^\beta e-\frac{m^2}{2}(e_{\alpha\beta}e^{\beta\alpha}+c\, e^2)+\nonumber \\ &\,&+d_1\, R\,e^{\alpha\beta}\,e_{\alpha\beta}+d_2\,R\,e^2+d_3\, R_{\alpha\beta\mu\nu}\,e^{\alpha\mu}\,e^{\beta\nu}+d_4\,R_{\alpha\beta}\,e^{\alpha\mu}\,{e^{\beta}}_\mu+d_5\, R_{\alpha\beta}\,e^{\alpha\beta}\,e+\nonumber \\ &\,&+d_6 \,R_{\alpha\beta\mu\nu}\,e^{\alpha\beta}\,e^{\mu\nu}+d_7\, R_{\alpha\beta}\,e^{\alpha\mu}\,{e_\mu}^\beta+d_8\,R \,e^{\alpha\beta}\, e_{\beta\alpha}+d_9 \,R_{\alpha\beta}\,e^{\mu\alpha}\,{e_\mu}^\beta \label{lcg} \end{aligned}\] where \(d_j\) \((j=1,2...,9)\) are arbitrary constants for now. The equations of motion are: \[\begin{aligned}
E_{\rho\sigma}\doteq \dfrac{\delta S_{\mbox{\tiny{nFP}}}^g(c)}{\delta e^{\rho\sigma}}&=&\frac{1}{2}\square (e_{\rho\sigma}+e_{\sigma\rho})+\frac{1}{6}\nabla_\rho\nabla^\lambda e_{\lambda\sigma}-\frac{1}{2}(\nabla_\rho\nabla^\lambda e_{\sigma\lambda}+\nabla_\sigma\nabla^\lambda e_{\rho\lambda})-\frac{1}{2}\nabla_\sigma \nabla^\alpha e_{\alpha\rho}+\nonumber \\ &\,&-\frac{1}{3}g_{\rho\sigma}\square e+\frac{1}{3}\nabla_\rho\nabla_\sigma e+\frac{1}{3}g_{\rho\sigma}\nabla^\beta\nabla^\alpha e_{\alpha\beta}-m^2(e_{\sigma\rho}+c\, e\, g_{\sigma\rho})+2d_1R\,e_{\rho\sigma}+\nonumber\\ &\,&+2d_2\, R\, e\, g_{\rho\sigma}+2d_3\, R_{\rho\beta\sigma\nu}\,e^{\beta\nu}+2d_4\,R_{\rho\beta}\, {e^{\beta}}_\sigma +d_5\, R_{\rho\sigma}\,e +d_5 \, g_{\rho\sigma}\, R^{\alpha\beta}\, e_{\alpha\beta}+\nonumber \\ &\,&+ 2d_6\, R_{\alpha\beta\rho\sigma}\, e^{\alpha\beta}+d_7\, {R^\alpha}_\sigma\, e_{\alpha\rho}+d_7\, {R_\rho}^\alpha e_{\sigma\alpha}+2d_8\, R \, e_{\sigma\rho}+2d_9\, {R_\sigma}^\beta\, e_{\rho\beta}=0\nonumber \\ \end{aligned}\] The vector constraint is the following expression: \[\begin{aligned}
\mathcal{C}_\rho\doteq\nabla^\sigma E_{\rho\sigma} &=&(1-2d_3-2d_6)R_{\rho\lambda\sigma\alpha}\nabla^\alpha e^{\lambda\sigma}+(1+2d_6)R_{\rho\lambda\sigma\alpha}\nabla^\alpha e^{\sigma\lambda}+d_5R_{\alpha\beta}\nabla_\rho e^{\alpha\beta}+\nonumber \\ &\,&+\biggl(\frac{1}{2}-2d_3-2d_6+d_7\biggr)e^{\lambda\sigma}\nabla_\lambda R_{\sigma\rho}+(-1+2d_3+d_5)e^{\lambda\sigma}\nabla_\rho R_{\sigma\lambda}+\nonumber \\ &\,&+\biggl(\frac{1}{2}+2d_4+2d_6\biggr)e^{\sigma\lambda}\nabla_\lambda R_{\sigma\rho}+\biggl(\frac{1}{2}+2d_9\biggr)R^{\lambda\mu}\nabla_\mu e_{\rho\lambda}+2d_1R\nabla^\sigma e_{\rho\sigma}+\nonumber \\ &\,&+\biggl(\frac{1}{2}+d_7\biggr)R^{\lambda\alpha}\nabla_\alpha e_{\lambda\rho}+ \biggl(\frac{1}{6}+d_7\biggr)R_{\lambda\rho}\nabla_\mu e^{\mu\lambda}+\biggl( \frac{1}{4}+2d_1+d_9\biggr)e_{\rho\sigma}\nabla^\sigma R+\nonumber \\ &\,& +\biggl(2d_4-\frac{1}{2}\biggr)R_{\lambda\rho}\nabla_\mu e^{\lambda\mu}+\biggl(\frac{1}{4}+\frac{d_7}{2}+2d_8\biggr)e_{\lambda\rho}\nabla^\lambda R+(2d_8R-m^2)\nabla^\sigma e_{\sigma\rho}+\nonumber \\ &\,&+\biggl(\frac{1}{3}+d_5\biggr)R_{\alpha\rho}\nabla^\alpha e +(2d_2R-m^2c)\nabla_\rho e+\biggl(2d_2+\frac{d_5}{2}\biggr) e\nabla_\rho R =0 \label{vvlc} \end{aligned}\] The tensor constraint will be obtained from the expression below: \[\begin{aligned}
\mathcal{C}_{\rho\sigma}\doteq E_{\rho\sigma}-E_{\sigma\rho} &=&\frac{2}{3}(\nabla_\rho\nabla^\lambda e_{\lambda\sigma}-\nabla_\sigma\nabla^\lambda e_{\lambda\rho})+\biggl(m^2+2(d_1-d_8)R\biggr)(e_{\rho\sigma}-e_{\sigma\rho})+\nonumber \\ &\,&+2(2d_6+d_3)R_{\rho\sigma\alpha\beta}\,e^{\alpha\beta}+(2d_4-d_7)({R_\rho}^\beta\, e_{\beta\sigma}-{R_\sigma}^\beta \, e_{\beta\rho})+\nonumber \\ &\,&+(d_7-2d_9)({R_\rho}^\beta\, e_{\sigma\beta}-{R_\sigma}^\beta \, e_{\rho\beta})=0 \label{vtlc} \end{aligned}\] Regarding the scalar constraint, due to the Weyl symmetry of the kinetic terms in \(\mathcal{L}^{g}_{\mbox{\tiny{nFP}}}(c)\) we do not need to add second derivatives of equations of motion in order to produce a constraint as in ([\[venovo\]](#venovo){reference-type="ref" reference="venovo"}). We can simply have: \[\begin{aligned}
\mathcal{C} &\doteq & g^{\rho\sigma}E_{\rho\sigma}\nonumber \\ &=&\biggl[-m^2(1+4c)+(2d_1+8d_2+d_5+2d_8)R\biggr]e+2(d_3+d_4+2d_5+d_7+d_9)R_{\sigma\beta}e^{\sigma\beta}=0\nonumber \\ &\,& \label{vec-2} \end{aligned}\] Thus, we have a scalar constraint in arbitrary gravitational backgrounds. In the appendix we show that although, it is possible to get \(e=0\) from ([\[vec-2\]](#vec-2){reference-type="ref" reference="vec-2"}) in arbitrary backgrounds, we are not able to have a curved space version of the tensor constraint without restricting the background space. Henceforth we assume Einstein spaces ([\[es\]](#es){reference-type="ref" reference="es"}). Let us rewrite ([\[vvlc\]](#vvlc){reference-type="ref" reference="vvlc"}), ([\[vtlc\]](#vtlc){reference-type="ref" reference="vtlc"}) and ([\[vec-2\]](#vec-2){reference-type="ref" reference="vec-2"}) as follows: \[\begin{aligned}
\mathcal{C}_\rho\doteq\nabla^{\sigma}E_{\rho\sigma}&=&(1-2d_3-2d_6)R_{\rho\lambda\sigma\mu}\nabla^{\mu}e^{\lambda\sigma}+(1+2d_6)R_{\rho\lambda\sigma\mu}\nabla^{\mu}e^{\sigma\lambda} +\tilde{d_1}R\, \nabla^\sigma e_{\rho\sigma}+\nonumber \\ &\,&+\Biggl[\biggl(\frac{1}{6}+2\tilde{d_8}\biggr)R-m^2\Biggr]\nabla^{\sigma}e_{\sigma\rho}+\Biggl[\biggl(\frac{1}{12}+2\tilde{d_2}\biggr)R-m^2c\Biggr]\nabla_{\rho}e=0 \label{vvc} \end{aligned}\] \[\begin{aligned}
\mathcal{C}_{\rho\sigma}\doteq E_{\rho\sigma}-E_{\sigma\rho}&=&\frac{2}{3}(\nabla_\rho\nabla^\lambda e_{\lambda\sigma}-\nabla_\sigma\nabla^\lambda e_{\lambda\rho})+2(d_3+2d_6)R_{\rho\sigma\alpha\beta}e^{\alpha\beta}+\nonumber \\ &\,& +\Biggl[m^2+\biggl(2\tilde{d_1}-2\tilde{d_8}+\frac{d_9}{2}\biggr)R\Biggr](e_{\rho\sigma}-e_{\sigma\rho})=0\label{vtc} \end{aligned}\] \[\begin{aligned}
\mathcal{C} &\doteq & g^{\rho\sigma}E_{\rho\sigma}=\biggl[-m^2(1+4c)+\biggl(2\tilde{d_1}+8\tilde{d_2}+\frac{d_3}{2}+2\tilde{d_8}\biggr)R\biggr]e=0 \label{vec} \end{aligned}\] where we have defined \[\begin{aligned}
\tilde{d_1}&\doteq &d_1+\dfrac{d_4}{4}+\dfrac{d_9}{4}\nonumber \\ \tilde{d_2}&\doteq &d_2+\dfrac{d_5}{4}\label{dtilde} \\ \tilde{d_8}&\doteq &d_8+\dfrac{d_7}{4}\nonumber \end{aligned}\] Therefore, ([\[vec\]](#vec){reference-type="ref" reference="vec"}) leads to the scalar constraint \(e=0\), provided the coefficient of \(e\) is different from zero. On the other hand, ([\[vtc\]](#vtc){reference-type="ref" reference="vtc"}) still has terms with second derivatives. In order to solve this, we need \(\nabla^\mu e_{\mu\nu}=0\). We can get this automatically from the vector constraint ([\[vvc\]](#vvc){reference-type="ref" reference="vvc"}) by setting: \[\begin{aligned}
\tilde{d_1}=0,\hspace{0.5cm} d_3=1, \hspace{0.5cm} d_6=-\frac{1}{2} \end{aligned}\] as far as the coefficient of \(\nabla^\mu e_{\mu\nu}\) does not vanish. Back in the tensor constraint ([\[vtc\]](#vtc){reference-type="ref" reference="vtc"}) we obtain \(e_{[\mu\nu]}=0\) as far as its coefficient is nonvanishing too. In summary, all 11 Fierz-Pauli constraints ([\[vinculosa1-1\]](#vinculosa1-1){reference-type="ref" reference="vinculosa1-1"}), ([\[vinculosa1-2\]](#vinculosa1-2){reference-type="ref" reference="vinculosa1-2"}) and ([\[vinculosa1-3\]](#vinculosa1-3){reference-type="ref" reference="vinculosa1-3"}) are confirmed if \[\begin{aligned}
\tilde{m}^2\biggl(\tilde{m}^2-\frac{R}{6}\biggr)\Biggl\{(1+4c)\tilde{m}^2+\biggl[8(c\, \tilde{d}_8-\tilde{d}_2)-\frac{1}{2}\biggr]R\Biggr\}\neq 0 \label{condmtil} \end{aligned}\] \[\begin{aligned}
\tilde{m}^2\equiv m^2-2\tilde{d}_8R \end{aligned}\] while the equations of motion become: \[\begin{aligned}
E_{\rho\sigma}=(\square-\tilde{m}^2)\, e_{\rho\sigma}+2R_{\rho\alpha\sigma\beta}\ e^{\alpha\beta} \end{aligned}\] where the free paramenters \(\tilde{d_2}\) and \(\tilde{d_8}\) must satisfy the conditions ([\[condmtil\]](#condmtil){reference-type="ref" reference="condmtil"}).
### Local symmetries of \(\mathcal{L}^{g}_{\mbox{\tiny{nFP}}}(c)\)
The model \(\mathcal{L}^{g}_{\mbox{\tiny{nFP}}}(c)\) also presents vector and scalar symmetries. There is one vector symmetry which comes from the transformation \[\begin{aligned}
\delta e_{\rho\sigma}=\nabla_\sigma A_\rho \end{aligned}\] where \(A_\rho\) is an arbitrary vector. We need a nonvanishing scalar curvature (\(R\neq 0\)) and the following conditions: \[\begin{aligned}
\tilde{d_1}&=&0; \hspace{0.5cm} d_6=-\frac{1}{2}; \hspace{0.5cm}d_3=1;\nonumber \\ \tilde{m}^2&=&\frac{R}{6}\\ \tilde{d_2}&=&-\frac{1}{24}+\frac{c}{12}+\tilde{d_8}c\nonumber \end{aligned}\] Such conditions imply that the vector constraint ([\[vvc\]](#vvc){reference-type="ref" reference="vvc"}) be identically null, i.e., \(\nabla^\sigma E_{\rho\sigma}\equiv 0\). On the other hand, starting from the general scalar transformation \[\begin{aligned}
\delta e_{\rho\sigma}=A_1 \nabla_\rho \nabla_\sigma \lambda +A_2 \, g_{\rho\sigma}\, \lambda +A_3\, g_{\rho\sigma}\, \square \, \lambda \end{aligned}\] where \(\lambda\) is an arbitrary field and \(A_j\) \((j=1,2,3)\) are arbitrary constants, we have found three scalar symmetries for \(\mathcal{L}^{g}_{\mbox{\tiny{nFP}}}(c)\). The first one is \[\begin{aligned}
{\delta e_{\rho\sigma}}^{\mbox{\tiny{(1)}}}=g_{\rho\sigma}\ \lambda \label{symm1} \end{aligned}\] where we must have \[\begin{aligned}
(1+4c)\tilde{m}^2=R\biggl[\frac{1}{2}+8(\tilde{d}_2-c\, \tilde{d}_8)\biggr] \label{cond_sim3} \end{aligned}\] The second scalar symmetry is \[\begin{aligned}
{\delta e_{\rho\sigma}}^{\mbox{\tiny{(2)}}}=-4\nabla_\rho\nabla_\sigma \lambda+g_{\rho\sigma}\, \square \lambda \label{symm2} \end{aligned}\] where the relations below must hold: \[\begin{aligned}
d_3&=&1\nonumber \\ \tilde{m}^2&=&(1+12\tilde{d}_1)\frac{R}{6} \label{cond_sim4} \end{aligned}\] Finally, the last scalar symmetry is given by: \[\begin{aligned}
{\delta e_{\rho\sigma}}^{\mbox{\tiny{(3)}}}=\nabla_\rho\nabla_\sigma\lambda \label{symm3} \end{aligned}\] where the conditions below must be obeyed: \[\begin{aligned}
d_3&=&1\nonumber \\ \tilde{d_2}&=&-\frac{1}{24}+\frac{c}{12}+c(\tilde{d_1}+\tilde{d_8})\label{condsymm3} \\ \tilde{m}^2&=&(1+12\tilde{d}_1)\frac{R}{6}\nonumber \label{cond_sim5} \end{aligned}\] The symmetry ([\[symm3\]](#symm3){reference-type="ref" reference="symm3"}) and conditions ([\[condsymm3\]](#condsymm3){reference-type="ref" reference="condsymm3"}) follow from the previous scalar symmetries ([\[symm1\]](#symm1){reference-type="ref" reference="symm1"}) and ([\[symm2\]](#symm2){reference-type="ref" reference="symm2"}).
# Conclusion {#concl}
Here we have studied massive spin-2 models via a nonsymmetric rank-2 tensor in a curved background. As in the usual Fierz-Pauli (FP) case with a symmetric tensor, nonminimal couplings are necessary. The work here is a preliminary one and parallels the work on the FP model. As in that case we have assumed that the Ansätze ([\[La1g\]](#La1g){reference-type="ref" reference="La1g"}) and ([\[lcg\]](#lcg){reference-type="ref" reference="lcg"}) are linear on curvatures and their coefficients are analytic functions of \(m^2\). Although there seems to be slightly more freedom now in choosing the background metric than in the FP case, we have selected, for simplicity, background spaces of the Einstein type as in. We have succeeded in finding nontrivial solutions for the coefficients of our Ansätze by getting rid of second derivatives in the tensor, vector and scalar constraints. In particular, we have generalized from maximally symmetric spaces to Einstein spaces a previous work in the literature carried out for a massive spin-2 theory with nonsymmetric tensor \(e_{\mu\nu} \ne e_{\nu\mu}\), which corresponds to the model \({\cal L}(a_1)\) of section 2 at the specific point \(a_1=-1/4\). Regarding the model \({\cal L}_{nFP}(c)\), due to the Weyl symmetry of the kinetic terms, the scalar constraint has easily led to the traceless condition \(e=0\) but now the problem moved to the tensor constraint and once again we have found convenient to choose Einstein spaces. Comparing the results obtained here in Einstein spaces for \({\cal L}(a_1)\) and \({\cal L}_{nFP}(c)\) with the ones obtained in for the usual FP model, the main difference is that, besides the scalar and vector constraint, we now have a tensor constraint \(C_{\rho\mu}=0\), see ([\[vt\]](#vt){reference-type="ref" reference="vt"}) and ([\[vtnovo\]](#vtnovo){reference-type="ref" reference="vtnovo"}). However, due to the vector constraint \(\nabla^{\mu}e_{\mu\nu} =0\), the constraint \(C_{\rho\mu}=0\) amounts to \(e_{\mu\nu}-e_{\nu\mu} =0\) without further restriction in the background. So there seems to be no fundamental difference to the usual FP case in curved space. Since in the FP case there is no restriction on the background metric when we allow the coefficients to be non-analytic functions of \(m^2\), we would like to address that point also in the case of our nonsymmetric models. This is under investigation now. Moreover, in both cases of \({\cal L}(a_1)\) and \({\cal L}_{nFP}(c)\), we believe that less restrictive conditions on the coefficients can be obtained by getting rid of second order derivatives of time only, this is under study. We are also analyzing the special points in the parameters space where the local symmetries mentioned in the previous sections show up. They indicate massless and partially massless theories even if \(m^2 \ne 0\). The truly massless cases \(m=0\) in both \({\cal L}(a_1)\) and \({\cal L}_{nFP}(c)\) theories are also worth investigating in curved space. Especially in the second case where, at least in the flat space, we have massless spin-2 particles just like in the massless version of the FP model. As a final comment, we notice that the ghost free massive gravity theories, see and, accommodate massive gravitons propagating in any gravitational background, see. Those results agree with earlier perturbative (in powers of \(1/m^2\) ) calculations. Thus, if we obtain correct Lagrangian constraints for the models discussed here, in the case of nonanalytic coefficients, we would be prompted to search for nonlinear (self-interacting) versions of \({\cal L}(a_1)\) and \({\cal L}_{nFP}(c)\).
# Appendix-Constraints in a general background
## \({\cal L}(a_1)\)
We are looking for a solution of the system (14)-(20) without imposing restrictions on the background space. Since \(f_i\) \((i=1..,9)\) and \(b_j\) \((j=0..,3)\) are constants, we demand that the coefficients of \(R\) in the equations ([\[systemLa1-1Ap\]](#systemLa1-1Ap){reference-type="ref" reference="systemLa1-1Ap"}) and ([\[systemLa1-2Ap\]](#systemLa1-2Ap){reference-type="ref" reference="systemLa1-2Ap"}) are null, i.e., \[\begin{aligned}
b_0\biggl(\frac{1}{2}+6a_1\biggr)+2b_2\biggl(\frac{1}{4}+a_1\biggr)+2b_3(f_1+f_8)=0\\ b_0\biggl(\frac{1}{2}+6a_1\biggr)+2b_2\biggl(\frac{1}{4}+a_1\biggr)-2b_3f_2=0 \quad. \end{aligned}\] The solution, singular at \(a_1 = 1/4\) as expected, is given by \[\begin{aligned}
f_3&=&1\nonumber \\ f_4&=&\frac{1}{2}+f_9\nonumber \\ f_5&=&\frac{(1+4a_1)}{(-1+4a_1)}\nonumber \\ f_7&=&-\frac{8a_1}{(-1+4a_1)}-2f_9\\ f_2&=&-\frac{1+8a_1(1+2a_1)}{4(-1+4a_1)}+\frac{b_0}{2b_1}\nonumber \\ f_1&=& \frac{1+8a_1(1+2a_1)}{4(-1+4a_1)}-\frac{b_0}{2b_1}-f_8\nonumber \\ b_3&=&\frac{1}{2}\, b_1(1+12a_1)\nonumber \\ b_2&=&-\frac{b_1(1+12a_1)(1+4a_1)}{2(-1+4a_1)}\nonumber \label{system}\end{aligned}\] Plugging this solution back in ([\[ve\]](#ve){reference-type="ref" reference="ve"}), we obtain a scalar contraint in a general background: \[\begin{aligned}
\mathcal{C}&=&\mathcal{C}_1=\frac{(1+12a_1)}{(-1+4a_1)}\Biggl\{-\frac{b_1(1+4a_1)^2}{2(-1+4a_1)}R^{\rho\sigma}R_{\rho\sigma} e+\frac{b_1(1-12a_1)}{2}\nabla^\lambda R^{\sigma\rho}\nabla_\rho e_{\lambda\sigma}+\nonumber \\ &\,&+4b_1a_1e_{\lambda\sigma}\square R^{\lambda\sigma}+\frac{b_1(1+12a_1)}{2(-1+4a_1)}\nabla_\rho R^{\lambda\sigma}\nabla^\rho e_{\lambda\sigma}-\frac{b_1(1+4a_1)}{2}\nabla^\sigma R^{\lambda\alpha}\nabla_\alpha e_{\lambda\sigma}\Biggr\}+\nonumber \\ &\,&+\frac{b_1(1+12a_1)}{2}\Biggl\{\biggl[-\frac{b_0}{b_1}+\frac{1+8a_1(1+2a_1)}{4(-1+4a_1)}\biggr]\nabla^\lambda R\, \nabla^\rho e_{\rho\lambda}+\nonumber \\ &\,&-\biggl[\frac{b_0}{b_1}+\frac{1-2a_1(-1+4a_1)}{-1+4a_1}-2f_9\biggr]\nabla^\lambda R\, \nabla^\rho e_{\lambda\rho}\Biggr\}+\nonumber \\ &\,&+\frac{(1+12a_1)}{(-1+4a_1)}\Biggl\{(b_0R+b_1m^2)+\frac{b_1(1+4a_1)}{2}\biggl[m^2+\frac{b_0R}{b_1}-\frac{[16a_1(1+a_1)+3]R}{2(-1+4a_1)}\biggr]\Biggr\}R^{\rho\sigma}e_{\rho\sigma}+\nonumber \\ &\,&+b_1(1+12a_1)R^{\rho\sigma}R_{\rho\beta\sigma\nu}e^{\beta\nu}+\frac{(1+12a_1)}{(-1+4a_1)}\biggl[\frac{b_1(1+12a_1)(1+4a_1)}{2(-1+4a_1)}-8b_1a_1\biggr]R^{\rho\sigma}R_{\rho\beta}\, {e^{\beta}}_\sigma +\nonumber \\ &\,&+\frac{b_1(1+12a_1)}{2}\Biggl\{\biggl[\frac{2b_0}{b_1}-\frac{1+16a_1(1+3a_1)}{4(-1+4a_1)}\biggr]\nabla_\lambda R \, \nabla^\lambda e+\biggl[\frac{b_0}{b_1}-\frac{2a_1(1+4a_1)}{(-1+4a_1)}\biggr] (\square R)\, e\Biggr\}+\nonumber \\ &\,&+\Biggl\{(b_0R+b_1m^2)\biggl[3m^2-\frac{1+16a_1(1+3a_1)}{2(-1+4a_1)}R+\frac{3b_0R}{b_1}\biggr]+\nonumber \\ &\,&-\frac{b_1(1+12a_1)(1+4a_1)}{2(-1+4a_1)}\biggr[m^2+\frac{b_0R}{b_1}-\frac{[1+8a_1(1+2a_1)]R}{2(-1+4a_1)}\biggr]R\Biggr\}e\label{82} \end{aligned}\] We have not been able to derive \(e=0\) from ([\[82\]](#82){reference-type="ref" reference="82"}) without restrictions on the background space.
## \({\cal L}_{nFP}\)
For arbitrary backgrounds if we choose \(d_9=-d_3-d_4-2d_5-d_7\) and \(d_2 =-(2d_1+d_5+2d_8)/8\) and assume \(c \ne-1/4\), the scalar constraint ([\[vec-2\]](#vec-2){reference-type="ref" reference="vec-2"}) becomes simply \(e=0\). Putting those results back in ([\[vvlc\]](#vvlc){reference-type="ref" reference="vvlc"}) and ([\[vtlc\]](#vtlc){reference-type="ref" reference="vtlc"}), we obtain: \[\begin{aligned}
\mathcal{C}_\rho &=& (1-2d_3-2d_6)R_{\rho\lambda\sigma\alpha}\nabla^\alpha e^{\lambda\sigma}+(1+2d_6)R_{\rho\lambda\sigma\alpha}\nabla^\alpha e^{\sigma\lambda}+d_5R_{\alpha\beta}\nabla_\rho e^{\alpha\beta}+\nonumber \\ &\,&+\biggl(\frac{1}{2}-2d_3-2d_6+d_7\biggr)e^{\lambda\sigma}\nabla_\lambda R_{\sigma\rho}+(-1+2d_3+d_5)e^{\lambda\sigma}\nabla_\rho R_{\sigma\lambda}+\nonumber \\ &\,&+\biggl(\frac{1}{2}+2d_4+2d_6\biggr)e^{\sigma\lambda}\nabla_\lambda R_{\sigma\rho}+\biggl(\frac{1}{2}-2d_3-2d_4-4d_5-2d_7\biggr)R^{\lambda\mu}\nabla_\mu e_{\rho\lambda}+\nonumber \\ &\,&+2d_1R\nabla^\sigma e_{\rho\sigma}+\biggl( \frac{1}{4}+2d_1-d_3-d_4-2d_5-d_7\biggr)e_{\rho\sigma}\nabla^\sigma R+\nonumber \\ &\,& +\biggl(2d_4-\frac{1}{2}\biggr)R_{\lambda\rho}\nabla_\mu e^{\lambda\mu}+\biggl(\frac{1}{4}+\frac{d_7}{2}+2d_8\biggr)e_{\lambda\rho}\nabla^\lambda R+(2d_8R-m^2)\nabla^\sigma e_{\sigma\rho}+\nonumber \\ &\,&+\biggl(\frac{1}{2}+d_7\biggr)R^{\lambda\alpha}\nabla_\alpha e_{\lambda\rho}+ \biggl(\frac{1}{6}+d_7\biggr)R_{\lambda\rho}\nabla_\mu e^{\mu\lambda} \label{vvlc-2} \end{aligned}\] \[\begin{aligned}
\mathcal{C}_{\rho\sigma}&=&\frac{2}{3}(\nabla_\rho\nabla^\lambda e_{\lambda\sigma}-\nabla_\sigma\nabla^\lambda e_{\lambda\rho})+\biggl(m^2+2(d_1-d_8)R\biggr)(e_{\rho\sigma}-e_{\sigma\rho})+\nonumber \\ &\,&+2(2d_6+d_3)R_{\rho\sigma\alpha\beta}\,e^{\alpha\beta}+(2d_4-d_7)({R_\rho}^\beta\, e_{\beta\sigma}-{R_\sigma}^\beta \, e_{\beta\rho})+\nonumber \\ &\,&+(2d_3+2d_4+4d_5+3d_7)({R_\rho}^\beta\, e_{\sigma\beta}-{R_\sigma}^\beta \, e_{\rho\beta})=0 \label{vtlc-2} \end{aligned}\] There are still second derivatives in \(\mathcal{C}_{\rho\sigma}\). We could think of determining \(\nabla^{\sigma}e_{\sigma\rho}\) as a function of the remaining terms in ([\[vvlc-2\]](#vvlc-2){reference-type="ref" reference="vvlc-2"}) at \(d_8=0\) and plugging it back in ([\[vtlc-2\]](#vtlc-2){reference-type="ref" reference="vtlc-2"}) which leads to: \[\begin{aligned}
\mathcal{C_{\rho\sigma}}&=&\frac{4}{3m^2}\Biggl[(1-2d_3-2d_6)R_{[\sigma\lambda\beta\mu}\nabla_{\rho]}\nabla^\mu e^{\lambda\beta}+(1+2d_6)R_{[\sigma\lambda\beta\mu}\nabla_{\rho]}\nabla^\mu e^{\beta\lambda}+\nonumber \\ &\,&+\biggl(\frac{1}{2}-2d_3-2d_4-4d_5-2d_7\biggr)R^{\lambda\mu}\nabla_{[\rho}\nabla_\mu e_{\sigma]\lambda}+\biggl(\frac{1}{2}+d_7\biggr)R^{\lambda\mu}\nabla_{[\rho}\nabla_\mu e_{\lambda\sigma]}+\nonumber \\ &\,& +\biggl(\frac{1}{6}+d_7\biggr)R_{\mu[\sigma}\nabla_{\rho]}\nabla_\lambda e^{\lambda\mu}+\biggl(-\frac{1}{2}+2d_4\biggr)R_{\mu[\sigma}\nabla_{\rho]}\nabla_\lambda e^{\mu\lambda}+d_5R_{\alpha\beta}\nabla_{[\rho}\nabla_{\sigma]} e^{\alpha\beta}+\nonumber \\ &\,&+2d_1R\nabla_{[\rho}\nabla^\beta e_{\sigma]\beta}+\nabla_{[\rho} \mathcal{F}_{\sigma]}\Biggr]+\biggl(m^2+2(d_1-d_8)R\biggr)(e_{\rho\sigma}-e_{\sigma\rho})+\nonumber \\ &\,& +2(2d_6+d_3)R_{\rho\sigma\alpha\beta}\,e^{\alpha\beta}+(2d_4-d_7)({R_\rho}^\beta\, e_{\beta\sigma}-{R_\sigma}^\beta \, e_{\beta\rho})+\nonumber \\ &\,&+(2d_3+2d_4+4d_5+3d_7)({R_\rho}^\beta\, e_{\sigma\beta}-{R_\sigma}^\beta \, e_{\rho\beta}) \label{lastf} \end{aligned}\] where \(\mathcal{F}_\alpha\) does not contain any derivative of \(e_{\rho\sigma}\) and is defined as follows: \[\begin{aligned}
\mathcal{F}_\alpha &\doteq &+\biggl(\frac{1}{2}-2d_3-2d_6+d_7\biggr)e^{\lambda\mu}\nabla_\lambda R_{\mu\alpha}+(-1+2d_3+d_5)e^{\lambda\mu}\nabla_\alpha R_{\mu\lambda}+\nonumber \\ &\,&+\biggl(\frac{1}{2}+2d_4+2d_6\biggr)e^{\mu\lambda}\nabla_\lambda R_{\mu\alpha}+\biggl(\frac{1}{4}+2d_1-d_3-d_4-2d_5-d_7\biggr)e_{\alpha\lambda}\nabla^\lambda R+\nonumber \\ &\,& +\biggl(\frac{1}{4}+\frac{d_7}{2}\biggr)e_{\lambda\alpha}\nabla^\lambda R \end{aligned}\] Unfortunately, without any restriction on the background space we have not been able to avoid second derivatives of \(e_{\mu\nu}\) in the tensor constraint ([\[lastf\]](#lastf){reference-type="ref" reference="lastf"}). Therefore, just like the \(\mathcal{L}^g(a_1)\) case, we are led to Einstein spaces \(R_{\mu\nu}=\frac{R}{4}g_{\mu\nu}\) once again. .8cm
[^1]: Throughout this work we use \(\eta_{\mu\nu}=(-,+,+,+)\), \(e_{(\alpha\beta)}=(e_{\alpha\beta}+e_{\beta\alpha})/2\) and \(e_{[\alpha\beta]}=(e_{\alpha\beta}-e_{\beta\alpha})/2\).
[^2]: We disregard nonanalytic functions of \(m^2\) and the term \(R_{\alpha\beta\mu\nu}\, e^{\alpha\mu}\, e^{\nu\beta}\) which is redundant due to the cyclic property \(R_{\mu\nu\alpha\beta}+R_{\mu\alpha\beta\nu}+R_{\mu\beta\nu\alpha}=0\).
[^3]: Altogether with Bianchi identities we have \(\nabla^\mu R_{\mu\nu\rho\sigma}=0\) and \(\nabla^\mu R=\partial^\mu R=0\).
[^4]: If \(a_1=\dfrac{1}{4}\), those terms would be eliminated, but this specific value for \(a_1\) represents the FP case and it's not of our interest here.
[^5]: The model \(\mathcal{L}_{\mbox{\tiny{nFP}}}(c)\) can be generalized to arbitrary \(D\geq 3\), see. | {'timestamp': '2017-06-07T02:06:50', 'yymm': '1706', 'arxiv_id': '1706.01770', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01770'} |
# Introduction and summary
Scattering amplitudes are central for understanding the structure of superstring theory. Their low-energy (\(\alpha'\)) expansion exhibits many deep mathematical structures. For instance, the four-graviton scattering amplitude of type II string theory in \(10-d\) space-time dimensions is expected to be invariant under the string U-duality group \(E_{d+1}(\mathbb{Z})\) order by order in \(\alpha'\) . This can be used together with supersymmetry to determine the lowest order derivative corrections of the form \(D^{2k}R^4\) arising from the four-graviton scattering amplitude as exact (generalized) automorphic forms of the moduli . As these automorphic forms are invariant under U-duality, they contain information about all orders of string perturbation theory and also non-perturbative effects. However, most results here are restricted to the four-graviton amplitude in type II in various dimensions and BPS-protected couplings associated with small automorphic representations. The automorphic forms have led to predictions of higher-genus string amplitudes and certain non-renormalisation theorems that have been confirmed by direct calculations . In a different direction, superstring amplitudes have been evaluated for many particles at low orders in string perturbation theory and the systematics of their \(\alpha'\)-expansion has been studied. At string tree level and for the scattering of \(N\) open or closed strings, there are amazing systematics associated with the theory of (single-valued) multiple polylogarithms and (single-valued) multiple zeta values . At string one-loop order, the corresponding generalisation appears to be that of (single-valued) elliptic polylogarithms and (single-valued) elliptic multiple zeta values that is currently under construction . 'Single-valued' here indicates a certain projection on the set of multiple zeta values that has to be applied in the closed superstring case . Other references on the relations of loop integrals to multiple zeta values include and for other work on the modular structure of string one-loop amplitudes see for example . In the present paper, we are interested in functions that arise in (or are related to) the \(\alpha'\)-expansion of closed superstring one-loop amplitudes. A one-loop amplitude is given by an integral over the modulus \(\tau\) of the world-sheet torus where the integrand is a modular \(SL(2,\mathbb{Z})\)-invariant function that is determined by world-sheet conformal field theory. The integrand depends on \(\alpha'\) and therefore the \(\alpha'\)-expansion of the one-loop amplitude can be studied from an \(\alpha'\)-expansion of the integrand. The separation into analytic and non-analytic terms in \(\alpha'\) can be effectively implemented by studying the behaviour of the integrand near the boundary of the torus moduli space (cutting off the \(\tau\) integration on the \(SL(2,\mathbb{Z})\) fundamental domain). This separation and the structure of this expansion was studied in where a formalism was developed that represented the integrand at a given \(\alpha'\)-order by a Feynman diagram of the world-sheet conformal field theory. This has led to the study of the structure and systematics of such world-sheet Feynman diagrams and the associated integrands in their own right . Understanding the structure of the integrand is necessary for finding the integrated value that is the actual contribution to the scattering amplitude. We note that for string amplitudes with more than four external states, the integrands are not necessarily described in terms of scalar propagators only but there can also be derivatives of propagators appearing in the world-sheet Feynman diagrams . In , it was shown that in the five-graviton case one can remove these derivatives up to order \(D^{10}R^4\). However, it is not known whether this is possible to arbitrary derivative order or for more general amplitudes. Restricting to standard scalar Feynman diagrams will therefore perhaps not capture all possible contributions to string scattering. Nevertheless, the scalar Feynman diagrams exhibit already a rich mathematical structure that is worthwhile to investigate. The integrand functions determined by the scalar world-sheet Feynman diagrams are now called *modular graph functions* and several cases have been studied in great detail. For world-sheet Feynman diagrams with one and two loops, the complete structure of the connected Feynman diagrams in terms of their behaviour under the modular Laplacian has been worked out and this has led to many interesting and unexpected identities among these modular graph functions that partially mirror identities of multiple polylogarithms . Beyond this complete treatment of one and two loops, some special cases of higher loop integrand functions have been analysed and some of them have been integrated . One of the main tools in the study of these functions are the modular invariant differential equations that they satisfy. These are typically inhomogeneous Laplace equations that sometimes admit an explicit integration with boundary conditions from degeneration limits of the toroidal world-sheet. In this paper, we will study an infinite family of modular graph functions at three-loop order on the world-sheet. We restrict to tetrahedral Feynman diagrams but allow for an arbitrary number of vertices along the edges of the tetrahedron. In graphical notation, the functions we are interested in are associated with Feynman diagrams of the form \[\begin{aligned}
\begin{tikzpicture} \point{2}{-1} \point{-2}{-1} \point{0}{2} \point{0}{0} \draw[thick] (2,-1)--(-2,-1)--(0,2)--(2,-1)--(0,0)--(0,2); \draw[thick] (-2,-1)--(0,0); \node [above] at (0,2) {\(\scriptstyle z_{1}\)}; \node [above right] at (0,0) {\(\scriptstyle z_{2}\)}; \node [below] at (-2,-1) {\(\scriptstyle z_{3}\)}; \node [below] at (2,-1) {\(\scriptstyle z_{4}\)}; \greens{-1}{0.5}{s} \greens{0}{1}{t} \greens{1}{0.5}{p} \greens{1}{-0.5}{q} \greens{-1}{-0.5}{w} \greens{0}{-1}{r} \end{tikzpicture}\nonumber \end{aligned}\] where the labels on the edges indicate the number of consecutive scalar propagators along the edge, meaning that the corresponding propagator is raised to the power given by the label. The simplest instance of such a modular graph function, corresponding to the case \(s=t=p=q=w=r=1\), was studied in and its contribution to the \(D^{12}R^4\) derivative correction was determined by using the inhomogeneous Laplace equation satisfied by the integrand associated with this diagram. In general, we will refer to the modular graph functions associated with the above diagram as tetrahedral modular graph functions. We will call \(s+t+p+q+w+r\) the *weight* of the modular graph function. We shall show in this paper that the family of tetrahedral modular graph functions satisfies an inhomogeneous Laplace equation where the right-hand side contains 'simpler' modular graph functions when the spectrum is diagonalised. This is in complete parallel with where at two loops the right-hand sides were quadratic polynomials in non-holomorphic Eisenstein series. Our results contain those of mentioned above as a special case and we employ heavily graphical methods similar to those of . The tetrahedral graph is symmetric under the action of the finite permutation group \(\mf{S}_4\) and we will show that the modular Laplace operator is closely related to the quadratic Casimir operator of \(\mf{sl}(3)\). These two ingredients allow us to use finite group theory and representation theory to deduce certain properties of the spectrum of the Laplace operator acting on the family of tetrahedral modular graph functions. As a by-product we will obtain a simple rederivation of the two-loop results of by the same methods. We note that the modular Laplacian on tetrahedral modular graph functions closes without the need to introduce modular graph *forms* that were recently introduced as a generalisation in . Contrary to the cases studied in , the eigenvalues of the modular Laplacian that we obtain are surprisingly not only of the form \(s(s-1)\) for non-negative integers \(s\). There are many possible generalisations and extensions of our work that are beyond the scope of the present paper. A point we have not investigated systematically is to use the inhomogeneous Laplace equations that we find to determine a basis of independent modular graph functions. This point would be very interesting in particular in connection with (elliptic) multiple polylogarithms. It would also be relevant for performing the actual world-sheet integrals over the modular graph functions that we do not attempt here. We note that useful techniques for determining the behaviour of the modular graph functions in the degeneration limit of the world-sheet (\(\tau_2\to\infty\)) can be found in . Finally, it would be interesting to consider extensions of the tetrahedral modular graph functions to also include derivatives in such a way that one reconstructs integrands of closed superstring one-loop amplitudes with five and more external legs. A widely open field is also the extension to higher genus string amplitudes, see for some relevant work on genus-two Riemann surfaces, in particular in connection with the so-called Kawazumi--Zhang invariant. The structure of this article is as follows. We first review in section [2](#sec:G1){reference-type="ref" reference="sec:G1"} general facts about genus-one amplitudes in closed superstring theory in order to motivate the types of Feynman diagrams and Laplace equations that we analyse. This includes an exposition of the diagrammatical tools for manipulating modular graphs. In section [3](#sec:TMF){reference-type="ref" reference="sec:TMF"}, we introduce the tetrahedral modular graph functions that are the central objects in this paper. We present their Laplace equation in general and introduce a generating function that makes it possible to connect to the representation theory of \(\mf{sl}(3)\). We present detailed examples of Laplace equations and spectral properties up to weight \(12\) together with some general considerations. These are the main results of this paper. Appendices contain results on simpler two-loop modular graph functions and technical details of some of the calculations of section [3](#sec:TMF){reference-type="ref" reference="sec:TMF"}.
# Genus-one amplitudes and modular graph functions {#sec:G1}
We shall consider genus-one contributions to \(n\)-graviton scattering amplitudes in type II string theory compactified on a torus \(T^{d}\) from ten to \(D=10-d\) space-time dimensions. The moduli dependence of these contributions appears generically through integrals of the type \[\label{eq:genI} \mathcal{I}_{F} = \int_{\mathcal{F}} d\mu \, F(\bar{\tau}, \tau) \, \Gamma_{(d,d)} \,.\] The integration domain \(\mathcal{F}\) is a fundamental domain of the moduli space, of genus-one Riemann surface \[\label{eq:FD} \mathcal{F} = \left\{ |\tau_1|\leq \frac12, | \tau | \geq 1 \right\} = \mathbb{H} / PSL_2(\mathbb{Z})\,,\] where \(\mathbb{H}=\left\{ \tau =\tau_1+i\tau_2 \in \mathbb{C} \, \middle|\, \tau_2>0\right\}\) is the complex upper half plane on which the modular group \(PSL(2,\mathbb{Z})\) acts by the standard fractional linear transformation. The integration measure \(d \mu\) is the standard \(PSL(2,\mathbb{Z})\) invariant measure \[d \mu = \frac{d \tau_1 d \tau_2}{ \tau_2^2} \,,\] such that the volume of the fundamental domain is normalised to be \(\int_{\mathcal{F}} d \mu = \frac{\pi}{3}\). The function \(\Gamma_{(d,d)}\) is the Narain genus-one partition function for the self-dual lattice that describes toroidal compactifications from ten dimensions to \(10-d\) dimensions: \[\label{eq:Narain} \Gamma_{(d,d)} = \tau_{2}^{d/2} \sum_{n^{I} \in \mathbb{Z}^{2d}} e^{-\pi \tau_2 | Z(n) |^2} e^{\pi i \tau_1 n^I \eta_{IJ} n^J}\,,\] in terms of the \(SO(d,d)\) invariant metric \(\eta_{IJ}=\left(\begin{smallmatrix} 0&1\\1&0\end{smallmatrix}\right)\). The mass squared \(| Z(n) |^2\) appearing in [\[eq:Narain\]](#eq:Narain){reference-type="eqref" reference="eq:Narain"} is given by \[|Z(n)|^{2} = Z_{a}(n) Z^{a}(n) = n^{I} \mathcal{V}_{a I} \mathcal{V}^{a}{}_{J} n^{J} \,.\] \(\mathcal{V}^{a}{}_{I}\) is the coset representative parametrising the symmetric space \(SO(d,d)/SO(d) \times SO(d)\), which transforms from the left under the local compact subgroup and from the right under the global \(SO(d,d)\). Local coordinates can be chosen in terms of the metric and \(B\)-field on the torus in the standard fashion . The Narain partition function \(\Gamma_{(d,d)}\) is invariant under \(SO(d,d,\mathbb{Z})\) transformations and modular transformations from \(PSL(2,\mathbb{Z})\). The invariance under \(SO(d,d,\mathbb{Z})\) is obvious, \(PSL(2,\mathbb{Z})\) invariance can only be seen after using Poisson resummation. The integral [\[eq:genI\]](#eq:genI){reference-type="eqref" reference="eq:genI"} is by construction still a function of the moduli \(\mathcal{V}^a{}_I\) of the space-time Narain torus; the dependence on the world-sheet torus parameter \(\tau\) is being integrated over. The function \(F(\tau,\bar{\tau})\) appearing in [\[eq:genI\]](#eq:genI){reference-type="eqref" reference="eq:genI"} encodes the specifics of the scattering process under consideration. It is required to be invariant under \(PSL(2, \mathbb{Z})\) transformations acting by \[F(\tau, \bar{\tau}) = F \biggl(\frac{ a\tau + b}{ c\tau + d}, \frac{ a\bar{\tau} + b}{c\bar{\tau} + d} \biggr) \,, \qquad a d-b c = 1 \,,\] with \(a,b,c,d\in\mathbb{Z}\). For general processes \(F(\tau,\bar{\tau})\) term will be a complicated function encoding momentum and Narain moduli dependence. However, for a four-graviton interaction, its form can be found explicitly as a Koba--Nielsen prefactor \[\label{eq:4pt} F(\tau,\bar{\tau}) = \prod_{i = 1}^{4} \int_{\Sigma} \frac{d^2 z_i}{\tau_2} e^{\mathcal{D}} \,,\] where \(\mathcal{D}\) is a sum over all insertions at local coordinates \(z_i\) and \(z_j\) \[\label{eq:defD} \mathcal{D} = \sum_{i < j} s_{i j} G(z_i-z_j | \tau) \,\] with dimensionless Mandelstam variables \[\label{eq:MS} s_{i j} =-\frac{1}{4} \alpha^{\prime} (k_i + k_j)^2\] and \(G(z_i-z_j|\tau)\) the translation invariant scalar propagator between \(z_i\) and \(z_j\) on the world-sheet torus of modulus \(\tau\). We will give an explicit form for the propagator below in [\[eq:Pmom\]](#eq:Pmom){reference-type="eqref" reference="eq:Pmom"}. The integral in [\[eq:4pt\]](#eq:4pt){reference-type="eqref" reference="eq:4pt"} is over the world-sheet of the torus \(\Sigma\) that we parametrise in a fixed domain of volume \(\int_\Sigma d^2z = \tau_2\), where \(d^2z = dz_1 dz_2\) in terms of the real and imaginary parts of \(z=z_1+iz_2\). While the formula [\[eq:4pt\]](#eq:4pt){reference-type="eqref" reference="eq:4pt"} is correct for four-graviton scattering, one will have additional insertions beyond Koba--Nielsen factors for higher point amplitudes and there can be additional Narain moduli dependences. The integral \(\mathcal{I}_F\) in [\[eq:genI\]](#eq:genI){reference-type="eqref" reference="eq:genI"} is an object of central interest in string theory. However, no closed formula for it is known. From a low-energy perspective, one can consider the \(\alpha'\)-expansion of the integral, corresponding to an expansion in \(s_{ij}\ll 1\). This generates analytical (in \(\alpha'\)) terms in the scattering amplitude . Sometimes one can then understand the integrand, and possibly even the integral, at a fixed order in \(\alpha'\) by studying the differential equations the integrand satisfies. The integral \(\mathcal{I}_F\) is a function on Narain moduli space \(SO(d,d)/SO(d)\times SO(d)\) and for the case when \(F(\tau,\bar{\tau})\) is independent of the Narain moduli one can compute the action of the \(SO(d,d)\) Laplacian by using \[\begin{aligned}
\left[\Delta_{SO(d,d)}-2 \Delta_{SL(2)} +\frac14d(d-2)\right] \Gamma_{(d,d)} = 0 \end{aligned}\] that relates the \(SO(d,d)\) action to one of the modular invariant \(SL(2)\) Laplace operator acting on \(\tau\). By partial integration the action of the \(SO(d,d)\) Laplacian can then be mapped to the \(SL(2)\) Laplacian acting on \(F(\tau,\bar{\tau})\). This action of \(\Delta_{SL(2)}\), more specifically in an \(\alpha'\) expansion, is what we shall study in this paper. For genus-one world sheets with metrics parametrised by \(\tau\in\mathcal{F}\) of [\[eq:FD\]](#eq:FD){reference-type="eqref" reference="eq:FD"} the \(SL(2)\) invariant Laplacian on the upper half plane is \[\Delta \equiv \Delta_{SL(2)} = 4 \tau_2^2 \partial_{\tau} \partial_{\bar{\tau}} = \tau_2^2 \left(\partial_{\tau_1}^2 + \partial_{\tau_2}^2\right)\] and we shall henceforth drop the subscript \(SL(2)\) on the Laplacian as it is the only one we will use.
## Low-order contributions to the four-graviton amplitude
From the definition of \(\mathcal{D}\) in [\[eq:defD\]](#eq:defD){reference-type="eqref" reference="eq:defD"} and the Mandelstam invariants we see that we can perform a low-energy expansion in the \(s_{ij}\) corresponding to the \(\alpha'\) expansion of string theory. The result of expanding the exponential in [\[eq:4pt\]](#eq:4pt){reference-type="eqref" reference="eq:4pt"} can be represented by a world-sheet Feynman diagram consisting of four points that are connected by \(n\) lines, where \(n\) is the order of the expansion in Mandelstam variables. The function \(F(\tau,\bar{\tau})\) controlling the four-graviton amplitude expands for small momenta \(s_{i j} \ll 1\) (corresponding to a power expansion in \(\alpha'\)) as \[\label{eq:Fexp} F(\tau,\bar{\tau}) = \sum_{n = 0}^{\infty} \frac{1}{n!} \prod_{i = 1}^{4} \int_{\Sigma} \frac{d^2 z_i}{\tau_2} \biggl( \sum_{j < k} s_{j k} G(z_j-z_k| \tau) \biggr)^n \,,\] where \(n\) counts the number of world-sheet propagators between the \(4\) points \(z_i\). By virtue of the definition [\[eq:MS\]](#eq:MS){reference-type="eqref" reference="eq:MS"} of the \(s_{ij}\) this corresponds also to the power of \(\alpha'\). Because of momentum conservation the \(s_{ij}\) are not all independent in the massless four-point amplitude. Letting \[\begin{aligned}
s= s_{12}=s_{34} \,,\quad t = s_{13}=s_{24} \,,\quad t=s_{23}=s_{14} \end{aligned}\] one has \[s + t + u = 0 \,.\] As a consequence, the analytic part of the four-graviton amplitude can be expanded in a double series in \(s^2+t^2+u^2\) and \(s^3+t^3+u^3\), except for the classical Einstein--Hilbert contribution . The low order terms in the expansion of \(F\) can be found for example in along with their integrals contributing to the four-graviton amplitude in the low-energy expansion. We recall that the one-loop string theory calculation in ten space-time dimensions gives contributions to \(R^4\) and then to every even derivative order starting from \(D^6R^4\). (In lower dimensions one also has contributions for \(D^4R^4\).) The integrated contributions up to \(D^{10}R^4\) have been worked out . A further discussion of the low order contributions can be found in .
## Laplacian on modular graph functions and the Green's function
For high order in \(\alpha'\) the explicit functional dependence of \(F\) on \(\tau\) is not very well understood. As in , one can consider the function [\[eq:Fexp\]](#eq:Fexp){reference-type="eqref" reference="eq:Fexp"} that generates the world-sheet diagrams in the \(\alpha'\)-expansion as a prototype of a new class of functions called *modular graph functions* that are constructed from world-sheet Feynman diagrams with an arbitrary number of points (not only four) connected by scalar propagators. These diagrams will not directly correspond to string processes but can serve as an interesting class of modular functions and certainly are relevant to the string theory calculation. In view of the structure of [\[eq:Fexp\]](#eq:Fexp){reference-type="eqref" reference="eq:Fexp"} that is given by an integral over products Green's function connecting different vertices we will follow and study the following more general class of functions where the integrand is given by \[\label{eq:IMG} \mathcal{I}(\{n_{kl}\},\tau,\bar{\tau}) = \prod_{i = 1}^{n} \int_{\Sigma} \frac{d^{2} z_{i}}{\tau_2} \biggl( \prod_{k < l} G^{n_{kl}}(z_{k}-z_{l} | \tau) \biggr) \,\] for some non-negative integer powers \(n_{kl}\) and a total of \(n\) vertices. Compared to [\[eq:Fexp\]](#eq:Fexp){reference-type="eqref" reference="eq:Fexp"} we have removed the Mandelstam variables and allowed for an arbitrary number \(n\) of vertices. Dimensionally, such an integrand would be related to an amplitude with low-energy action of the form \(D^{2w} R^4\), where \(w=\sum_{k<l} n_{kl}\) is the weight of the integrand. However, this is generally only true dimensionally as it is known that for \(n>4\) the integrand of the genus one amplitude is not of this simple form but also involves derivatives of Green's functions . Functions arising from modular integrals over expressions of the type [\[eq:IMG\]](#eq:IMG){reference-type="eqref" reference="eq:IMG"} are called *modular graph functions*. As we will review below, they can be represented graphically in terms of Feynman world-sheet diagrams. Modular graph functions are invariant under modular transformations acting on \(\tau\) and can appear as constituents of higher derivative corrections. Besides this physical relevance, they represent an interesting new class of modular functions on the upper half plane and we will be interested in evaluating the modular Laplacian acting on them. We shall use a diagrammatic way of computing the Laplacian acting on a modular graph function, similar to . The basic tool is to first rewrite the modular Laplacian using deformation theory as \[\begin{aligned}
\label{eq:Ldef} \Delta = 4 \tau_2^2 \partial_\tau \partial_{\bar{\tau}} = \delta_\mu \bar{\delta}_\mu\,, \end{aligned}\] where \(\delta_\mu\) denotes infinitesimal changes in the complex structure while keeping the coordinates fixed. The advantage of this formalism is that one can work out the deformation of the Green's function on general grounds. As shown in , the action of the deformation \(\delta_\mu\) acting on a single Green's function connecting two points \(z_i\) and \(z_j\) can be replaced by the insertion of an additional vertex: \[\label{eq:deltaG} \delta_\mu G( z_{i}-z_{j}|\tau) =-\frac{1}{\pi} \int_{\Sigma} d^{2} z \partial_{z} G(z-z_{i}|\tau) \partial_{z} G({z-z_{j}|\tau}) \,.\] Moreover, the Laplacian with respect to the modular parameter satisfies \[\Delta G( z_{i}-z_{j}|\tau) = 0 \,,\] since the result becomes a total derivative. Therefore, when evaluating the modular Laplacian on a product of Green's functions one has to apply the deformations \(\delta_\mu\) and \(\bar{\delta}_\mu\) to different factors. Obviously the function \(\mathcal{I}(\tau,\bar{\tau})\) has in general highly complicated dependence on \(\tau, \bar{\tau}\), however some very simple and elegant answers were found in the past. For the purpose of understanding these results we can express \(G\) through a lattice sum: \[\label{eq:Pmom} G(z| \tau) = \sum_{(m,n) \in \mathbb{Z}^2\setminus\{(0,0)\}} \frac{\tau_{2}}{\pi |m\tau + n|^{2}} e^{\frac{\pi}{\tau_2} (\bar{z}(m \tau + n)-z(m \bar{\tau} + n))} \,,\] where the integers \(m, n\) parametrise the discrete momenta on a torus \(p = m + n \tau\). The zero momentum was removed from the lattice sum \(\mathbb{Z}^{2}\). In this representations the modular invariance of \(G(z| \tau)\) can be easily seen. The scalar Green's functions on a world-sheet torus \(\Sigma\) of modulus \(\tau\) satisfies the identities \[\label{eq:G_id} \partial_{z} \partial_{\bar{z}} G(z | \tau) =-\pi \delta^{(2)}(z) + \frac{\pi}{\tau_2} \,, \qquad \int_{\Sigma} d^{2} z G(z| \tau) = 0 \,,\] where the second condition is related to the choice of zero mode. The zero mode does not contribute to [\[eq:Fexp\]](#eq:Fexp){reference-type="eqref" reference="eq:Fexp"} by momentum conservation \(\sum_{i<j} s_{ij}=0\). It is very helpful to represent the Green's functions and their derivatives in a graphical way in order to simplify the calculations. These are the modular graphs that represent the integrands of the modular integrals. A single point is a symbol for the integration of an insertion and lines represent the Green's functions between two insertions \[\begin{tikzpicture} \filldraw[black] (-7,-0.5) circle (2pt) node[anchor=north] {\(i\)} node[anchor=east]{\(\frac{1}{\tau_2} \int_{\Sigma} d^2 z_{i} = \;\)}; \draw[gray, thick] (-1,-0.5)--(1,-0.5); \filldraw[black] (-1,-0.5) circle (2pt) node[anchor=north] {\(i\)} node[anchor=east]{\(\hspace{-4 mm}\,, \, \hspace{4 mm}\frac{1}{\tau_2^2} \int_{\Sigma} d^2 z_{i} d^2 z_{j} \, G(z_{i}-z_{j}|\tau) \; = \; \)}; \filldraw[black] (1,-0.5) circle (2pt) node[anchor=north] {\(j\)}; \node at (1.3,-0.5) {\(.\)}; \end{tikzpicture}\] Due to reflection invariance in \(z\) one does not need to put arrows on the propagators. In the same way several lines joining at one point \(z\) mean that several Green's functions connect an insertion point \(z\) to various other insertions and all of them are integrated out at the end \[\begin{tikzpicture} \draw[gray, thick] (-1,-0.5)--(1,0.5); \draw[gray, thick] (-1,-0.5)--(1,0); \draw[gray, thick] (-1,-0.5)--(1,-0.5); \draw[gray, thick] (-1,-0.5)--(1,-1); \draw[gray, thick] (-1,-0.5)--(1,-1.5); \filldraw[black] (-1,-0.5) circle (2pt) node[anchor=north] {\(z\)} node[anchor=east]{\(\frac{1}{\tau_2^{n + 1}} \int_{\Sigma} d^2 z \prod_{i = 1}^{n} \int_\Sigma d^2 z_{i} \, G(z-z_{i}|\tau) \; = \; \)}; \filldraw[black] (1,0.5) circle (2pt) node[anchor=north] {\(\scriptstyle n\)}; \filldraw[black] (1,0) circle (2pt) node[anchor=north] {\(\scriptstyle n-1\)}; \filldraw[black] (1,-0.5) circle (2pt) node[anchor=north] {\(\scriptstyle... \qquad.\)}; \filldraw[black] (1,-1) circle (2pt) node[anchor=north] {\(\scriptstyle 2\)}; \filldraw[black] (1,-1.5) circle (2pt) node[anchor=north] {\(\scriptstyle 1\)}; \end{tikzpicture}\] We also introduce the action of the derivative in \(z\) acting on one of the Green's functions \[\begin{tikzpicture} \draw[gray, thick] (-1,-0.5)--(1,-0.5); \filldraw[black] (-1,-0.5) circle (2pt) node[anchor=north] {\(i\)} node[anchor=south west] {\(\,\,\partial\)} node[anchor=east]{\(\frac{1}{\tau_2^2} \int_{\Sigma} d^2 z_{i} d^2 z_{j} \, \partial_{z_{i}} G(z_{i}-z_{j}|\tau) \; = \; \)}; \filldraw[black] (1,-0.5) circle (2pt) node[anchor=north] {\(j\)}; \node at (-0.2,-1.5) {\(=-\)}; \draw[gray, thick] (0.3,-1.5)--(2.3,-1.5); \filldraw[black] (0.3,-1.5) circle (2pt) node[anchor=north] {\(i\)}; \filldraw[black] (2.3,-1.5) circle (2pt) node[anchor=north] {\(j\)} node[anchor=south east] {\(\,\,\partial\)} node[anchor=west]{\( \; = \;-\frac{1}{\tau_2^2} \int_{\Sigma} d^2 z_{i} d^2 z_{j} \, \partial_{z_{j}} G(z_{i}-z_{j}|\tau) \,.\)}; \end{tikzpicture}\] Here, we have also illustrated the consequence of translation invariance \(\partial_{z_i} G(z_i-z_j|\tau) =-\partial_{z_j} G(z_i-z_j|\tau)\). Last but not least we are always able to rewrite the action of the deformation \(\delta_\mu\) into the action of derivatives on the world-sheet by introducing an additional insertion, such that \[\label{eq:deltaGDiag} \begin{tikzpicture} \draw[gray, thick] (-1,-0.5)--(1,-0.5); \filldraw[black] (-1,-0.5) circle (2pt) node[anchor=north] {\(i\)} node[anchor=south west] {\(\qquad \delta_\mu\)} node[anchor=east]{\(\delta_\mu \biggl[ \frac{1}{\tau_2^2} \int_{\Sigma} d^2 z_{i} d^2 z_{j} G(z_{i}-z_{j}|\tau) \biggr] \; = \; \)}; \filldraw[black] (1,-0.5) circle (2pt) node[anchor=north] {\(j\)} node[anchor=west] {\(\; = \;\)}; \draw[gray, thick] (-2.5,-1.5)--(-0.5,-1.5); \filldraw[black] (-2.5,-1.5) circle (2pt) node[anchor=north] {\(i\)} node[anchor=east] {\(\quad-\frac{\tau_2}{\pi} \)}; \filldraw[black] (-1.5,-1.5) circle (2pt) node[anchor=north] {\(z\)} node[anchor= south east] {\(\partial \,\)} node[anchor= south west] {\(\, \partial\)}; \filldraw[black] (-0.5,-1.5) circle (2pt) node[anchor=north] {\(j\)} node[anchor=west]{\(\; = \;-\frac{1}{\pi \tau_2^2} \int_{\Sigma} d^{2} z_{i} d^{2} z_{j} d^{2} z \partial_{z} G(z-z_{i}|\tau) \partial_{z} G({z-z_{j}|\tau}) \,. \)}; \end{tikzpicture}\] This rule is due to [\[eq:deltaG\]](#eq:deltaG){reference-type="eqref" reference="eq:deltaG"}. There is a similar formula for the conjugate deformation \(\bar{\delta}_\mu\) in terms of the conjugate world-sheet derivative \(\bar{\partial}_z\). Because of equation [\[eq:G_id\]](#eq:G_id){reference-type="eqref" reference="eq:G_id"} we see that every diagram with at least one node that has only one Green's function connecting to it is vanishing. Therefore for tadpoles diagrams we obtain \[\label{eq:tadpole} \begin{tikzpicture} \draw[gray, thick, dashed] (-2,-0.5)--(0,0); \draw[gray, thick, dashed] (-2,0)--(0,0); \draw[gray, thick] (0,0)--(2,0); \draw[gray, thick, dashed] (-2,0.5)--(0,0); \filldraw[black] (0,0) circle (2pt) node[anchor=north] {\(i\)}; \filldraw[black] (2,0) circle (2pt) node[anchor=north] {\(z\)} node[anchor=west] {\(\; \propto \; \frac{1}{\tau_2} \int_{\Sigma} d^{2} z G(z_{i}-z| \tau) = 0 \,.\)}; \end{tikzpicture}\] Additionally, we read out from [\[eq:G_id\]](#eq:G_id){reference-type="eqref" reference="eq:G_id"} the diagrammatical simplification rule \[\label{eq:ddcontract} \begin{tikzpicture} \draw[gray, thick, dashed] (-3,-0.5)--(-2,0); \draw[gray, thick, dashed] (-3,0)--(-2,0); \draw[gray, thick, dashed] (-3,0.5)--(-2,0); \draw[gray, thick, dashed] (0,0)--(1,-0.5); \draw[gray, thick, dashed] (0,0)--(1,0); \draw[gray, thick, dashed] (0,0)--(1,0.5); \draw[gray, thick] (-2,0)--(0,0); \filldraw[black] (-2,0) circle (2pt) node[anchor=north] {\(i\)} node[anchor=south west] {\(\partial\)}; \filldraw[black] (0,0) circle (2pt) node[anchor=north] {\(j\)} node[anchor=south east] {\(\bar{\partial}\)}; \filldraw[black] (1.5,0) circle (0pt) node[anchor=center] {\(\; = \frac{\pi}{\tau_2} \)}; \draw[black,thick] (2.1,-0.6)--(2.0,-0.6)--(2.0,0.6)--(2.1,0.6); \draw[black,thick] (4.1,-0.6)--(4.2,-0.6)--(4.2,0.6)--(4.1,0.6); \draw[gray, thick, dashed] (2.1,-0.5)--(3.1,0); \draw[gray, thick, dashed] (2.1,0)--(3.1,0); \draw[gray, thick, dashed] (2.1,0.5)--(3.1,0); \draw[gray, thick, dashed] (3.1,0)--(4.1,-0.5); \draw[gray, thick, dashed] (3.1,0)--(4.1,0); \draw[gray, thick, dashed] (3.1,0)--(4.1,0.5); \filldraw[black] (3.1,0) circle (2pt) node[anchor=north] {\(i = j\)}; \draw[gray, thick, dashed] (5.2,-0.5)--(6.2,0); \draw[gray, thick, dashed] (5.2,0)--(6.2,0); \draw[gray, thick, dashed] (5.2,0.5)--(6.2,0); \draw[gray, thick, dashed] (7.2,0)--(8.2,-0.5); \draw[gray, thick, dashed] (7.2,0)--(8.2,0); \draw[gray, thick, dashed] (7.2,0)--(8.2,0.5); \draw[black,thick] (5.2,-0.6)--(5.1,-0.6)--(5.1,0.6)--(5.2,0.6); \draw[black,thick] (8.2,-0.6)--(8.3,-0.6)--(8.3,0.6)--(8.2,0.6); \filldraw[black] (4.7,0) circle (0pt) node[anchor=center] {\(\;-\frac{\pi}{\tau_2} \)}; \filldraw[black] (6.2,0) circle (2pt) node[anchor=north] {\(i\)}; \filldraw[black] (7.2,0) circle (2pt) node[anchor=north] {\(j\)}; \node at (8.4,0) {\(.\)}; \end{tikzpicture}\] The derivative with respect to world-sheet variables acting on one of the Green's functions can be moved on the graph reproducing the integration by parts formula. We obtain for example \[\begin{tikzpicture} \filldraw[black] (0,0) circle (2pt) node[anchor=north] {\(i\)} node[anchor=south east] {\(\partial\)}; \draw[gray, thick] (-1,0)--(0,0); \draw[gray, thick] (0,0)--(1,0.5); \draw[gray, thick] (0,0)--(1,0); \draw[gray, thick] (0,0)--(1,-0.5); \node at (1.5,0) {\(= \, \,-\)}; \filldraw[black] (3,0) circle (2pt) node[anchor=north] {\(i\)}; \draw[gray, thick] (2,0)--(3,0); \draw[gray, thick] (3,0)--(4,0.5); \draw[gray, thick] (3,0)--(4,0); \draw[gray, thick] (3,0)--(4,-0.5); \node at (3.2,0.35) {\(\partial\)}; \node at (4.5,0) {\(-\)}; \filldraw[black] (6,0) circle (2pt) node[anchor=north] {\(i\)}; \draw[gray, thick] (5,0)--(6,0); \draw[gray, thick] (6,0)--(7,0.5); \draw[gray, thick] (6,0)--(7,0); \draw[gray, thick] (6,0)--(7,-0.5); \node at (6.5,0.2) {\(\partial\)}; \node at (7.5,0) {\(-\)}; \filldraw[black] (9,0) circle (2pt) node[anchor=north] {\(i\)}; \draw[gray, thick] (8,0)--(9,0); \draw[gray, thick] (9,0)--(10,0.5); \draw[gray, thick] (9,0)--(10,0); \draw[gray, thick] (9,0)--(10,-0.5); \node at (9.25,-0.4) {\(\partial\)}; \node at (10.5,0) {\(\,.\)}; \end{tikzpicture}\]
## Modular graph functions with one and two world-sheet loops
Some subsets of modular functions with a particular geometric structures are well understood. For example a simple \(s\)-polygon of Green's functions reproduces non-holomorphic Eisenstein series with a somewhat unusual, yet for our purpose useful, normalisation : \[\begin{tikzpicture} \coordinate (A1) at (0,0); \coordinate (A2) at (0.75,0.5); \coordinate (A3) at (1.5,0); \coordinate (A4) at (1.5,-0.75); \coordinate (A5) at (0.75,-1.25); \coordinate (An) at (0,-0.75); \draw[gray, thick] (An)--(A1)--(A2)--(A3)--(A4)--(A5); \draw [dashed] (A5)--(An); \node [above left] at (0.45,0.15) {\(\scriptstyle 1\)}; \node [above right] at (1, 0.2) {\(\scriptstyle 2\)}; \node [right] at (1.45,-0.35) {\(\scriptstyle 3\)}; \node [right] at (1,-1.1) {\(\scriptstyle 4\)}; \node [left] at (0,-0.4) {\(\scriptstyle s\)}; \filldraw[black] (A1) circle (2pt); \filldraw[black] (A2) circle (2pt); \filldraw[black] (A3) circle (2pt); \filldraw[black] (A4) circle (2pt); \filldraw[black] (A5) circle (2pt); \filldraw[black] (An) circle (2pt); \node [right] at (1.7,-0.75/2) {\(\displaystyle \, = \, \sum_{(m, n) \in \mathbb{Z}^2/\{0\}} \frac{\tau_2^{s}}{\pi^{2 s} |m + n \tau|^{2 s}} = E_{s}(\tau, \bar{\tau}) \,.\)}; \end{tikzpicture}\] This is the simplest non-trivial structure that appears as the modular graph function, with a single summation over the discretised momentum in the loop and it depends just on a single value \(s\), that is the number of internal vertex insertions. Eisenstein functions are know to satisfy a homogeneous Laplace equation \[( \Delta-s(s-1) ) E_{s}(\tau,\bar{\tau}) = 0 \,.\] This equations can also be proved diagrammatically using the rules outlined above. For the particular value of \(s = 0\), we use the normalisation \[E_{0}(\tau,\bar{\tau}) = 1 \,.\] The next more complex structure was discussed in detail in and depends on a triplet \((s,t,p)\) of integer values, that describe the number of vertex insertions on the path connecting points \(z_1\) and \(z_2\) on the torus \[\label{eq:Cstp} \begin{tikzpicture} \coordinate (A1) at (0,-0.35); \coordinate (A11) at (0.75,-0.35); \coordinate (A12) at (1.75,-0.35); \coordinate (A13) at (2.75,-0.35); \coordinate (A14) at (3.75,-0.35); \coordinate (A2) at (0.75,0.5); \coordinate (A21) at (1.75,0.5); \coordinate (A22) at (2.75,0.5); \coordinate (A23) at (3.75,0.5); \coordinate (A24) at (4.5,-0.35); \coordinate (A3) at (4,0.5); \coordinate (An) at (0.75,-1.25); \coordinate (An1) at (1.75,-1.25); \coordinate (An2) at (2.75,-1.25); \coordinate (An3) at (3.75,-1.25); \draw[gray, thick] (A1)--(A11)--(A12); \draw[gray, thick] (A1)--(A2)--(A21); \draw[gray, thick] (A22)--(A23)--(A24); \draw[gray, thick] (A1)--(An)--(An1); \draw[gray, thick] (An2)--(An3)--(A24); \draw[gray, thick] (A13)--(A14)--(A24); \draw[gray, thick, dashed] (A21)--(A22); \draw[gray, thick, dashed] (An1)--(An2); \draw [dashed] (A12)--(A13); \node [left] at (A1) {\(\scriptstyle z_1\)}; \node [above] at (0.3,0.1) {\(\scriptstyle 1\)}; \node [above] at (1.25,0.5) {\(\scriptstyle 2\)}; \node [above] at (3.25,0.5) {\(\scriptstyle s-1\)}; \node [above] at (4.25,0.1) {\(\scriptstyle s\)}; \node [above] at (0.45,-0.4) {\(\scriptstyle 1'\)}; \node [above] at (1.25,-0.4) {\(\scriptstyle 2'\)}; \node [above] at (3.3,-0.4) {\(\scriptstyle (t-1)'\)}; \node [above] at (4.1,-0.35) {\(\scriptstyle t'\)}; \node [below] at (0.2,-0.7) {\(\scriptstyle 1''\)}; \node [below] at (1.25,-1.2) {\(\scriptstyle 2''\)}; \node [below] at (3.25,-1.2){\(\scriptstyle (p-1)''\)}; \node [below right] at (4.1,-0.6) {\(\scriptstyle p''\)}; \node [right] at (A24) {\(\scriptstyle z_2\)}; \node [right] at (A24) {\(\quad = \sum^{'}_{\substack{(m_1,n_1)\\(m_2,n_2)}} \frac{\tau_2^{s + t + p}}{\pi^{s + t + p} |m_1 + n_1 \tau|^{2 s} |m_2 + n_2 \tau|^{2 t} | m_1 + m_2 + (n_1 + n_2)\tau|^{2 p} }\)}; \node at (6.3,-1.5) {\(= C_{(s,t,p)}(\tau,\bar{\tau}) \,, \)}; \filldraw[black] (A1) circle (2pt); \filldraw[black] (A2) circle (2pt); \filldraw[black] (A21) circle (2pt); \filldraw[black] (A22) circle (2pt); \filldraw[black] (A23) circle (2pt); \filldraw[black] (A24) circle (2pt); \filldraw[black] (A11) circle (2pt); \filldraw[black] (A12) circle (2pt); \filldraw[black] (A13) circle (2pt); \filldraw[black] (A14) circle (2pt); \filldraw[black] (A5) circle (2pt); \filldraw[black] (An) circle (2pt); \filldraw[black] (An1) circle (2pt); \filldraw[black] (An2) circle (2pt); \filldraw[black] (An3) circle (2pt); \end{tikzpicture}\] where we sum over discrete momenta \(p_1\) and \(p_2\) in the loop, excluding zero and have solved overall momentum conservation. The prime on the sum indicates that we have to exclude all zero momentum propagators, *i.e.*, \((m_1,n_1)\neq (0,0)\), \((m_2,n_2)\neq (0,0)\) and \((m_1+m_2,n_1+n_2)\neq (0,0)\). In string theory only non-negative integer values for \(s\), \(t\) and \(p\) arise but as argued in the function \(C_{(s,t,p)}\) can be analytically continued to arbitrary complex values of the parameters. We will often suppress the arguments \(\tau\) and \(\bar\tau\). Starting now we will use following abbreviations to indicate the number of Green's functions that connect two points in a simple manner \[\begin{tikzpicture} \coordinate (A1) at (0,0); \coordinate (A2) at (1,0); \coordinate (A3) at (2,0); \coordinate (A4) at (3,0); \coordinate (A5) at (4,0); \coordinate (A6) at (5,0); \coordinate (A8) at (7,0); \filldraw[black] (A1) circle (2pt); \filldraw[black] (A2) circle (2pt); \filldraw[black] (A3) circle (2pt); \filldraw[black] (A4) circle (2pt); \filldraw[black] (A5) circle (2pt); \node [below] at (0.5,0) {\(\scriptstyle 1\)}; \node [below] at (1.5,0) {\(\scriptstyle 2\)}; \node [below] at (3.5,0) {\(\scriptstyle s \)}; \node [right] at (A5) {\(\, \, = \)}; \draw[gray, thick] (A1)--(A3); \draw[gray, thick] (A4)--(A5); \draw[gray, thick, dashed] (A3)--(A4); \draw[gray, thick,] (A6)--(A8); \filldraw[black] (A6) circle (2pt); \filldraw[black] (A8) circle (2pt); \node at (7.5,0) {\(.\)}; \greens{6}{0}{s} \end{tikzpicture}\] With this notation Eisenstein functions and \(C_{(s,t,p)}\) can be written in a more graphical way \[\begin{tikzpicture} \node [left] at (0,0) {\(\scriptstyle z_1\)}; \filldraw [black] (0,0) circle (2pt); \draw [black, thick] (0.6,0) circle (0.6 cm); \greens{1.2}{0}{s} \node [right] at (1.2,0) {\(\,\,\,\, = E_{s}(\tau,\bar{\tau}) \, ,\)}; \begin{scope}[shift={(1,0)}] \draw [thick] (4,0).. controls (5,1.35).. (6,0); \draw [thick] (4,0).. controls (5,0).. (6,0); \draw [thick] (4,0).. controls (5,-1.35).. (6,0); \node [left] at (4,0) {\(\scriptstyle z_1\)}; \node [right] at (6,0) {\(\scriptstyle z_2\)}; \greens{5}{1}{s} \greens{5}{0}{t} \greens{5}{-1}{p} \point{4}{0} \point{6}{0} \node [right] at (6.2,0) {\(\,\,\,\, = C_{\three{\hspace{-2mm}s}{\hspace{-2mm}t}{\hspace{-2mm}p}}(\tau,\bar{\tau}) \,.\)}; \end{scope} \end{tikzpicture}\] It is obvious that \(C_{(s,t,p)}\) is completely symmetric under permutations of the \((s,t,p)\) triplet. Furthermore for specific values of \((s,t,p)\) the function \(C_{(s,t,p)}\) simplifies to a quadratic polynomial in Eisenstein series \[\label{eq:CabcSpec} C_{\three{\hspace{-0mm}s}{\hspace{-0mm}t}{\hspace{-0mm}0}} = E_{s} E_{t}-E_{s + t} \,, \qquad C_{\three{\hspace{-0mm}s}{\hspace{-0mm}t}{\hspace{-0mm}\(\bf-\)1}} = E_{s-1} E_{t} + E_{s} E_{t-1} \,.\] This simplification can be easily seen from the lattice sum representation. Unfortunately, the differential equation satisfied by a general \(C_{(s,t,p)}\) function is not any more homogeneous and can be derived to be \[\begin{aligned}
\label{eq:LCabc} ( \Delta-\omega) C_{\three{\hspace{-0mm}s}{\hspace{-0mm}t}{\hspace{-0mm}p}} &= st \left( C_{\three{\hspace{-0mm}s +1}{\hspace{-0mm}t-1}{\hspace{-0mm}p}} + C_{\three{\hspace{-0mm}s-1}{\hspace{-0mm}t + 1}{\hspace{-0mm}p}} + C_{\three{\hspace{-0mm}s +1}{\hspace{-0mm}t + 1}{\hspace{-0mm}p-2}} -2 C_{\three{\hspace{-0mm}s +1}{\hspace{-0mm}t}{\hspace{-0mm}p-1}} -2 C_{\three{\hspace{-0mm}s}{\hspace{-0mm}t + 1}{\hspace{-0mm}p-1}}\right)\nonumber\\ & \quad +\textrm{ the two other pairs of lines }\,, \end{aligned}\] where the eigenvalue \(\omega\) is given by \[\omega = s(s-1) + t(t-1) + p(p-1) \,.\] The spectrum of the modular Laplacian was studied in great detail in . In appendix [\[app:2loop\]](#app:2loop){reference-type="ref" reference="app:2loop"}, we present a simple rederivation of the results of that paper based on an application of Molien's theorem combined with some representation theory of \(\mf{sl}(2)\).
# Tetrahedral family of modular graph functions {#sec:TMF}
In this section, we introduce the family of modular graph functions associated with the tetrahedral graph and an arbitrary number of propagators on all edges. We determine the inhomogeneous Laplace equation satisfied by such functions and study some degeneration limits. The spectrum of the Laplace operator on tetrahedral modular graph functions is partially determined using generating function techniques.
## Inhomogeneous Laplace equation for tetrahedral modular graphs
The next very symmetrical topology with three-valent vertices after the one above is that of a tetrahedron (or Mercedes graph): \[\begin{tikzpicture} \point{2}{-1} \point{-2}{-1} \point{0}{2} \point{0}{0} \draw[thick] (2,-1)--(-2,-1)--(0,2)--(2,-1)--(0,0)--(0,2); \draw[thick] (-2,-1)--(0,0); \node [above] at (0,2) {\(\scriptstyle z_{1}\)}; \node [above right] at (0,0) {\(\scriptstyle z_{2}\)}; \node [below] at (-2,-1) {\(\scriptstyle z_{3}\)}; \node [below] at (2,-1) {\(\scriptstyle z_{4}\)}; \greens{-1}{0.5}{s} \greens{0}{1}{t} \greens{1}{0.5}{p} \greens{1}{-0.5}{q} \greens{-1}{-0.5}{w} \greens{0}{-1}{r} \node [right] at (1,0.5) {\( \quad = \sum^{'}_{p_i} \frac{\tau_2^{s + t+ p + q + w +r}}{\pi^{s + t + p + q + w + r} |p_1 |^{2 s} |p_2 |^{2 t} |p_1 + p_2|^{2 p} |p_3|^{2 q} |p_1 + p_2 + p_3|^{2 r} |p_2 + p_3|^{2 w}} = C_{\six{\hspace{-3 mm}s}{t}{p}{q}{\hspace{-1mm }w}{r}} \)}; \end{tikzpicture}\] The restriction on the sum means that the discrete momenta \(p_i=m_i+n_i \tau\) for integers \(m_i,n_i\in \mathbb{Z}\) are all non-zero and similarly for all other propagators, *i.e.*, \(p_1+p_2\neq 0\), \(p_2+p_3\neq 0\) and \(p_1+p_2+p_3\neq 0\). We have already solved momentum conservation in the above expression and the loop momenta are labelled as \[\raisebox{-0.3\height}{\begin{tikzpicture}[thick,decoration={ markings, mark=at position 0.5 with {\arrow{>}}} ] \draw[postaction=decorate] (0,0)--(0,2); \draw[postaction=decorate] (0,0)--(1.73,-1); \draw[postaction=decorate] (-1.73,-1)--(0,0); \draw[postaction=decorate] (-1.73,-1)--(0,2); \draw[postaction=decorate] (0,0)--(0,2); \draw[postaction=decorate] (1.73,-1)--(-1.73,-1); \draw[postaction=decorate] (0,2)--(1.73,-1); \draw (-1.2,0.8) node {\(p_1\)}; \draw (-0.3,0.7) node {\(p_2\)}; \draw (1.6,0.7) node {\(p_1+p_2\)}; \draw (0.7,-0.1) node {\(p_3\)}; \draw (0,-1.3) node {\(p_1+p_2+p_3\)}; \draw (-0.2,-0.6) node {\(p_2+p_3\)}; \end{tikzpicture}}\] As is well-known, the tetrahedron has point symmetry group \(\mf{S}_4\) acting on it. An explicit form of the action of this symmetric group on the graph can be found for example in . For a tetrahedral modular graph function \(C\!\!\raisebox{-0.3\height}{\six{s}{t}{p}{q}{w}{r}}\) we will call \(s+t+p+q+r+w\) the *weight* of the function. The genuine first non-trivial case arises at weight \(6\) and was treated already in . We will re-derive it within our more general analysis below. Without solving momentum conservation the tetrahedral modular graph function \(C\!\!\raisebox{-0.3\height}{\six{s}{t}{p}{q}{w}{r}}\) can be expressed in the symmetric way through six lattice sums and four Kronecker deltas preserving momentum conservation at each vertex \(z_{i}\): \[C_{\six{\hspace{-2 mm}1}{2}{3}{4}{6}{5}} = \sideset{}{'}\sum_{(m_{i},n_{i}) \in \mathbb{Z}^{2}} \delta^{z_i}_{m 0} \delta^{z_{i}}_{n 0} \prod_{j = 1}^{6} \frac{\tau^{s_j}_{2}}{\pi^{s_j} |m_{j} + n_{j} \tau|^{2 s_j}}\,,\] where the labels \(1,\ldots,6\) on the left-hand side stand for the parameters \(s_1,\ldots,s_6\) appearing on the right-hand side. Using either graphical methods or the sum representation, we can evaluate the modular Laplacian on these tetrahedral modular graph functions to be \[\begin{aligned}
\label{eq:LC6} \bigl(\Delta-\omega \bigr) C_{\six{\hspace{-2 mm}s}{t}{p}{q}{w}{r}} &=st \Bigg( C_{\six{\hspace{-2 mm}\(\bf -\)1}{1}{}{}{}{}} +C_{\six{1}{\hspace{-2 mm}\(\bf -\)1}{}{}{}{}} +C_{\six{1}{1}{\hspace{-2 mm}\(\bf -\)2}{}{}{}} -2C_{\six{1}{}{\hspace{-2 mm}\(\bf -\)1}{}{}{}} -2C_{\six{}{1}{\hspace{-2 mm}\(\bf -\)1}{}{}{}} \Bigg)\nonumber\\ &\hspace{10mm}\textrm{+ 11 other adjacent pairs of lines}\nonumber\\ &\quad+sq\Bigg( C_{\six{1}{\hspace{-2 mm}\(\bf -\)2}{}{1}{}{}} +C_{\six{1}{}{\hspace{-2 mm}\(\bf -\)2}{1}{}{}} +C_{\six{1}{}{}{1}{\hspace{-2 mm}\(\bf -\)2}{}} +C_{\six{1}{}{}{1}{}{\hspace{-2 mm}\(\bf -\)2}} +2 C_{\six{1}{}{\hspace{-2 mm}\(\bf -\)1}{1}{\hspace{-2 mm}\(\bf -\)1}{}} +2 C_{\six{1}{\hspace{-2 mm}\(\bf -\)1}{}{1}{}{\hspace{-2 mm}\(\bf -\)1}} \nonumber\\ &\hspace{15mm} -2C_{\six{1}{}{\hspace{-2 mm}\(\bf -\)1}{1}{}{\hspace{-2 mm}\(\bf -\)1}} -2C_{\six{1}{\hspace{-2 mm}\(\bf -\)1}{\hspace{-2 mm}\(\bf -\)1}{1}{}{}} -2C_{\six{1}{}{\hspace{-2 mm}\(\bf -\)1}{1}{}{\hspace{-2 mm}\(\bf -\)1}} -2C_{\six{1}{}{}{1}{\hspace{-2 mm}\(\bf -\)1}{\hspace{-2 mm}\(\bf -\)1}} \Bigg) \nonumber\\ &\hspace{10mm}\textrm{+ 2 other opposite pairs of lines} \end{aligned}\] with the 'eigenvalue' \(\omega\) being \[\label{eq:om} \omega = s (s-1) + p (p-1) + q (q-1) + r (r-1) + t (t-1) + w (w-1)-2(w p + q s + r t)\] The final mixed term does not arise for the 'sunset' functions of and is formed as the sum over the three sets of non-adjacent (opposite) lines in the diagram. The notation in [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"} means that the indices on the corresponding lines of the diagrams are increased or decreased in the indicated places while maintaining the labels on the left-hand side of the equation. Thus \[\begin{aligned}
\label{eq:6short} \raisebox{-0.5\height}{\scalebox{1.5}{\six{\hspace{-2 mm}\(\bf -\)1}{}{}{}{1}{}}} \hspace{2mm}\equiv \raisebox{-0.5\height}{\scalebox{1.5}{\six{\hspace{-3 mm}s\(\bf-\)1}{t}{p}{q}{w\!\(\bf +\)\!1}{r}}} \end{aligned}\] We present some details on the derivation of [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"} in appendix [\[app:LC6\]](#app:LC6){reference-type="ref" reference="app:LC6"}.
## Degeneration limits
Below we will also require some degeneration limits of the tetrahedral graphs when some of the vertices come together. These are \[\begin{tikzpicture} \point{1}{-1} \point{-1}{-1} \point{0}{1} \draw[thick] (-1,-1)--(1,-1); \draw [thick] (-1,-1).. controls (-1.4,0).. (0,1); \draw [thick] (-1,-1).. controls (-0.6,0).. (0,1); \draw [thick] (1,-1).. controls (1.4,0).. (0,1); \draw [thick] (1,-1).. controls (0.6,0).. (0,1); \node [above] at (0,1) {\(\scriptstyle z_{1}\)}; \node [below] at (-1,-1) {\(\scriptstyle z_{2}\)}; \node [below] at (1,-1) {\(\scriptstyle z_{3}\)}; \greens{-1.2}{0}{s} \greens{-0.5}{0}{t} \greens{0.5}{0}{p} \greens{1.2}{0}{q} \greens{0}{-1}{r} \node [right] at (1,0) {\(\quad \,\,\, = \sum^{'}_{p_i} \frac{\tau_2^{s + t + p + q + r}}{\pi^{s + t + p + q + r} |p_1 |^{2 s} |p_2 |^{2 t} |p_3|^{2 p} |p_1 + p_2|^{2 r} |p_1 + p_2 + p_3|^{2 q} } = C_{\five{\hspace{-2mm}s}{\hspace{-2mm}t}{\hspace{-2mm}p}{\hspace{-2mm}q}{r}}(\tau,\bar{\tau})\,, \)}; \end{tikzpicture}\] \[\begin{tikzpicture} \point{0}{-1} \point{0}{1} \draw [thick] (0,-1).. controls (-1.4,0).. (0,1); \draw [thick] (0,-1).. controls (-0.6,0).. (0,1); \draw [thick] (0,-1).. controls (1.4,0).. (0,1); \draw [thick] (0,-1).. controls (0.6,0).. (0,1); \node [above] at (0,1) {\(\scriptstyle z_{1}\)}; \node [below] at (0,-1) {\(\scriptstyle z_{2}\)}; \greens{-1.1}{0}{s} \greens{-0.5}{0}{t} \greens{0.5}{0}{p} \greens{1.1}{0}{q} \node [right] at (1,0) {\(\quad \,\,\, = \sum^{'}_{p_i} \frac{\tau_2^{s + t + p +q}}{\pi^{s + t + p +q} |p_1 |^{2 s} |p_2 |^{2 t} |p_3|^{2 p} |p_1 + p_2 + p_3|^{2 q} } = C_{\four{\hspace{-3 mm} s}{\hspace{-3 mm} t}{\hspace{-3 mm} p}{\hspace{-3 mm} q}}(\tau,\bar{\tau})\)}; \end{tikzpicture}\]
## Laplace equations at low weights
We now evaluate explicitly [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"} for \(C\!\!\raisebox{-0.3\height}{\six{s}{t}{p}{q}{w}{r}}\) for low weights \(s+t+p+q+r+w\) starting from weight \(6\).
### Laplace equation at weight \(6\)
In order to illustrate the use of these equations, we re-derive the Laplace equation for the simplest non-trivial tetrahedral modular graph function that appears for weight \(6\). From [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"} one finds \[(\Delta + 6 )C_{\six{1}{1}{1}{1}{1}{1}} = 12 \, C_{\six{\(-\)1}{1}{1}{1}{2}{2}} + 12 \, C_{\six{\(-\)1}{1}{2}{1}{2}{1}}-24 C_{\six{0}{1}{1}{1}{2}{1}}-24 \, C_{\six{0}{0}{2}{1}{2}{1}} + 12 \, C_{\six{0}{1}{2}{0}{2}{1}} \,.\] We simplify the right hand-side of the equation using equations [\[eq:simpT\]](#eq:simpT){reference-type="eqref" reference="eq:simpT"}:
Putting the results together we obtain \[\begin{aligned}
\label{eq:L111111} (\Delta + 6 )C_{\six{1}{1}{1}{1}{1}{1}} = 48 C_{\three{1}{2}{3}}-12 E_3^2 + 12 E_6 \,. \end{aligned}\] This equation was derived in this form first in and is relevant for determining the \(D^{12}R^4\) correction at one loop.
### Laplace equation at weight \(7\)
At weight \(7\) there is only a single genuine tetrahedral modular graph function associated with the diagram \[\begin{aligned}
\resizebox{.1\textwidth}{!}{ \begin{tikzpicture} \draw [line width=1.5mm, gray, opacity = 0.3] (2,-1)--(-2,-1)--(0,2)--(2,-1)--(0,0)--(0,2); \draw [line width=1.5mm, gray, opacity = 0.3] (-2,-1)--(0,0); \greensB{-1}{0.5}{2} \greensB{0}{1}{1} \greensB{1}{0.5}{1} \greensB{1}{-0.5}{1} \greensB{-1}{-0.5}{1} \greensB{0}{-1}{1} \end{tikzpicture}} \end{aligned}\] Plugging this into the equation [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"}, we find in a first instance \[\begin{aligned}
(\Delta + 6 )C_{\six{2}{1}{1}{1}{1}{1}} &= 8 C_{\six{\(-\)1}{1}{1}{1}{3}{2}} + 8 C_{\six{\(-\)1}{1}{2}{1}{3}{1}} + 2 C_{\six{\(-\)1}{1}{1}{2}{2}{2}} + 2 C_{\six{\(-\)1}{1}{2}{2}{2}{1}} + 8 C_{\six{\(-\)1}{1}{2}{1}{2}{2}}-8 C_{\six{0}{0}{2}{1}{3}{1}}\nonumber\\ &\quad-8 C_{\six{0}{0}{2}{1}{2}{2}}-8 C_{\six{0}{0}{3}{1}{2}{1}}-8 C_{\six{0}{1}{1}{1}{3}{1}}-10 C_{\six{0}{1}{1}{1}{2}{2}} -4 C_{\six{0}{1}{1}{2}{2}{1}} + 8 C_{\six{0}{1}{2}{0}{3}{1}} \nonumber\\ &\quad + 4 C_{\six{0}{1}{2}{0}{2}{2}} -6 C_{\six{0}{1}{2}{1}{2}{1}}-4 C_{\six{0}{1}{2}{1}{1}{2}} \,. \end{aligned}\] For the simplifications we use again [\[eq:simpT\]](#eq:simpT){reference-type="eqref" reference="eq:simpT"} and there are many cancellations. Combining all the terms together the final Laplace equation at weight \(7\) is \[\begin{aligned}
(\Delta + 6 )C_{\six{2}{1}{1}{1}{1}{1}} =-2 C_{\five{1}{2}{1}{2}{1}} + 30 C_{\three{1}{2}{4}} + 18 C_{\three{1}{3}{3}} + 8 C_{\three{2}{2}{3}}-12 E_3 E_4 + 12 E_7\. \end{aligned}\]
## Generating function, its Laplace equation and \(\mf{sl}(3)\)
For understanding more general properties of the spectrum of the Laplacian on tetrahedral modular graph functions, it is useful to consider a generating function, similar to the one introduced in . For the tetrahedral graphs considered here we write it as \[\begin{aligned}
{\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{t{\scriptscriptstyle_2}}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}} &= \sum_{s,t,p,q,w,r = 1}^{\infty} t^{s-1}_1 t^{t-1}_2 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6 C_{\six{\hspace{-2 mm}s}{t}{p}{q}{w}{r}} \,. \end{aligned}\] In terms of the lattice sum this can be thought of as considering massive propagators between the vertices \(z_i\) \[{\mathcal{W}}(t_i, \tau, \bar{\tau}) = \sideset{}{'}\sum_{(m_{i},n_{i}) \in \mathbb{Z}^{2}} \delta^{z_i}_{m 0} \delta^{z_{i}}_{n 0} \prod_{j = 1}^{6} \frac{\tau_{2}}{\pi |m_{j} + n_{j} \tau|^{2}-t_r \tau_2}\.\] We will now determine the action of the Laplace operator on \({\mathcal{W}}\) from the Laplace equation [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"}. We begin with the 'eigenvalue' \(\omega\) shown in [\[eq:om\]](#eq:om){reference-type="eqref" reference="eq:om"}. The left-hand side of the Poisson equations can be expressed using the relation \[\sum_{s,t,p,q,w,r = 1}^{\infty} t^{s-1}_1 t^{t-1}_2 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6 s(s-1) C_{\six{\hspace{-2 mm}s}{t}{p}{q}{w}{r}} = t_1 \partial^2_{1} (t_1 {\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{t{\scriptscriptstyle_2}}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}} ) \,.\] Here we use the notation \(\partial_i\equiv \partial/\partial_{t_i}\) as a short-hand. This part can be rewritten for all legs as \(\left(\sum_{i = 1}^{6} t_i \partial^2_{i} t_i \right){\mathcal{W}}\). The mixed terms in \(\omega\) of [\[eq:om\]](#eq:om){reference-type="eqref" reference="eq:om"} can be written in terms of \(t\)-derivatives as \[\sum_{s,t,p,q,w,r = 1}^{\infty} t^{s-1}_1 t^{t-1}_2 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6 w p C_{\six{\hspace{-2 mm}s}{t}{p}{q}{w}{r}} = \partial_{3} \partial_{6} (t_3 t_6 {\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{t{\scriptscriptstyle_2}}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}} ) \,.\] We therefore deduce that \[\begin{aligned}
\label{eq:omgen} &\sum_{s,t,p,q,w,r = 1}^{\infty} t^{s-1}_1 t^{t-1}_2 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6 \omega C_{\six{s}{t}{p}{q}{w}{r}} =\left(\sum_{i=1}^6 t_i \partial_i^2 t_i-2 \sum_{i=1}^3 \partial_i \partial_{i+3} t_i t_{i+3}\right) {\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{t{\scriptscriptstyle_2}}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}} \nonumber\\ &\hspace{20mm}= \left(\mathfrak{D}^2-\mathfrak{D}-6-2 \sum_{<i,j>} t_i t_j \partial_i \partial_j-4 \sum_{i=1}^3 t_i t_{i+3} \partial_i \partial_{i+3} \right) {\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{t{\scriptscriptstyle_2}}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}} \,, \end{aligned}\] where \(\omega\) on the left-hand side is given by [\[eq:om\]](#eq:om){reference-type="eqref" reference="eq:om"}. We have separated the sum over pairs of edges into the \(12\) adjacent pairs \(<i,j>\) and the three opposite pairs \((i,i+3)\) for \(i=1,2,3\). The differential operator \[\begin{aligned}
\mathfrak{D} = \sum_{i=1}^6 t_i \partial_i \end{aligned}\] measures the degree of homogeneous polynomials in the \(t_i\). Next we analyse the inhomogeneous terms on the right-hand side of the Laplace equation [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"}. We will use again the short-hand [\[eq:6short\]](#eq:6short){reference-type="eqref" reference="eq:6short"} to indicate a number of propagators different from the standard one in \(C\!\!\raisebox{-0.3\height}{\six{s}{t}{p}{q}{w}{r}}\). As a rule of thumb, any shifted index will be associated with a shifted power on the corresponding variable \(t_i\) in the generating function. Shifting the summation back to the standard range will produce 'boundary terms' where some of the edges have the special values that also appear in [\[eq:simpT\]](#eq:simpT){reference-type="eqref" reference="eq:simpT"}. Generally, only the edges with negative shifts will contribute to these boundary terms; the positive \(+1\) shifts are innocuous as they only appear for the edges whose power also multiplies the corresponding contribution. Let us consider as an example the first term on the right-hand side of [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"} that contains an adjacent pair of lines: \[\begin{aligned}
\label{eq:RHSgen} &\sum_{s,t,p,q,w,r = 1}^{\infty} t^{s-1}_1 t^{t-1}_2 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6 st C_{\six{1}{\hspace{-2 mm}\(\bf -\)1}{}{}{}{}} = \partial_1 \partial_2\!\!\left[\! t_1 t_2 \sum_{\substack{p,q,r,w = 1\\tau = 0 \\ s= 2}} t^{s-2}_1 t^{t}_2 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6C_{\six{}{}{}{}{}{}}\right] \nonumber\\ &\hspace{20mm}= \partial_1 \partial_2 \left[t_2^2 \sum_{\substack{s,p,q,r,w = 1\\tau = 0 }} t^{s-1}_1 t^{t-1}_2 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6C_{\six{}{}{}{}{}{}} \right]\nonumber\\ &\hspace{20mm}= \partial_1 \partial_2 \left[t_2^2 {\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{t{\scriptscriptstyle_2}}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}} + t_2 {\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{0}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}}\right]\. \end{aligned}\] The last term comes from the \(t=0\) term in the sum and we have introduced the notation \[\begin{aligned}
{\mathcal{W}}_{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{0}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}} = \sum_{s,p,q,r,w = 1} t^{s-1}_1 t^{p-1}_3 t^{q-1}_4 t^{r-1}_5 t^{w-1}_6C_{\six{}{0}{}{}{}{}} ={\mathcal{W}}_{\six{\hspace{-1 mm}1}{2}{3}{4}{6}{5}} \Big|_{2\to0}\ \end{aligned}\] for the generating function of degenerate tetrahedral graph functions. The function \(C\!\!\raisebox{-0.3\height}{\six{}{0}{}{}{}{}}\) could in principle be simplified using [\[eq:simpT\]](#eq:simpT){reference-type="eqref" reference="eq:simpT"}, but it is more compact to leave it in this form. We see that a term on the right-hand side of [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"} contributes both to a differential operator acting on \({\mathcal{W}}\!\!\raisebox{-0.3\height}{\six{\hspace{-2 mm}t{\scriptscriptstyle_1}}{t{\scriptscriptstyle_2}}{t{\scriptscriptstyle_3}}{t{\scriptscriptstyle_4}}{t{\scriptscriptstyle_6}}{t{\scriptscriptstyle_5}}}\) and to degenerate boundary terms. Manipulations similar to [\[eq:RHSgen\]](#eq:RHSgen){reference-type="eqref" reference="eq:RHSgen"} can be performed for all the adjacent lines \(\langle i,j\rangle\) and opposite lines in [\[eq:LC6\]](#eq:LC6){reference-type="eqref" reference="eq:LC6"}. Summing up all the contributions then gives \[\begin{aligned}
\label{eq:LG6} \bigl( \Delta-\mathfrak{D}^2 + \mathfrak{D} + 6\bigr) {\mathcal{W}}_{\six{\hspace{-1 mm}1}{2}{3}{4}{6}{5}} &= \sum_{V_{ijk}} \bigl( t^2_i + t^2_j + t^2_k-2 t_i t_j-2 t_i t_k-2 t_j t_k \bigr) (\partial_{j} \partial_{k} + \partial_{i} \partial_{j} + \partial_{i} \partial_{k}) {\mathcal{W}}_{\six{\hspace{-1 mm}1}{2}{3}{4}{6}{5}} \nonumber\\ &\quad\quad+ \sum_{i = 1}^{3} ( (t_{i+1}-t_{i+2} + t_{i+4}-t_{i + 5})^2-4 t_{i} t_{i + 3} )\partial_{i} \partial_{i+3} {\mathcal{W}}_{\six{\hspace{-1 mm}1}{2}{3}{4}{6}{5}} + {\mathcal{R}} \,, \end{aligned}\] where the two sums arise from the adjacent lines \(\langle i,j\rangle\) coming together at a vertex \(V_{ijk}\) and the three pairs of opposite lines. We have also moved some of the terms in [\[eq:omgen\]](#eq:omgen){reference-type="eqref" reference="eq:omgen"} to the right. The term \({\mathcal{R}}\) contains all the contributions from degenerate diagrams and is given explicitly by \[\begin{aligned}
{\mathcal{R}}&=-\sum_{i=1}^6 \sum_{V_{pqi}} \left( \partial_p + \partial_q +(2t_p + 2t_q-t_i) \partial_p \partial_q\right) {\mathcal{W}}_{\six{\hspace{-1 mm}1}{2}{3}{4}{6}{5}} \Big|_{i\to0}\nonumber\\ &\quad+ \sum_{i=1}^6 \sum_{\substack{p=1\\p\notin\{i,i+3\}}}^3 \left(t_i +2t_{i+3}-2t_{p+1}-2t_{p+4}\right)\partial_p \partial_{p+3} {\mathcal{W}}_{\six{\hspace{-1 mm}1}{2}{3}{4}{6}{5}} \Big|_{i\to0}\nonumber\\ &\quad+ \sum_{i=1}^6 \sum_{\Delta_{ipq}\neq \Delta_{irs}}^6 \left(\partial_p +\partial_q\right)\left(\partial_r+\partial_s\right) {\mathcal{W}}_{\six{\hspace{-1 mm}1}{2}{3}{4}{6}{5}} \Big|_{i\to-1}\nonumber\\ &\quad+\!\bigg[ 2 \partial_{1} \partial_{4} \Bigl( {\mathcal{W}}_{\six{\hspace{-2 mm}1}{0}{3}{4}{6}{0}} +{\mathcal{W}}_{\six{\hspace{-2 mm}1}{2}{0}{4}{0}{5}}-{\mathcal{W}}_{\six{\hspace{-2 mm}1}{0}{0}{4}{6}{5}}-{\mathcal{W}}_{\six{\hspace{-2 mm}1}{0}{3}{4}{0}{5}}-{\mathcal{W}}_{\six{\hspace{-2 mm}1}{2}{0}{4}{6}{0}}-{\mathcal{W}}_{\six{\hspace{-2 mm}1}{2}{3}{4}{0}{0}} \Bigr) \nonumber\\ &\hspace{20mm}+ \textrm{two other pairs of opposite lines} \biggr] \end{aligned}\] (Indices are to be understood modulo \(6\).) The four terms are almost simpler to describe in words: The first term is a sum over the six edges \(i\) with \(p\) and \(q\) joining line \(i\) at a vertex; so for \(i=1\) it would be the two cases \((p,q)\in \{ (2,3), (5,6)\}\) because edge \(4\) is opposite of edge \(1\). The second term is also an outer sum over the edges \(i\) and the inner sum produces the two pairs of opposite edges not containing \(i\); for \(i=1\) it would be \((t_1+2t_4-2t_3-2t_6)\partial_2\partial_5+ (t_1+2t_4-2t_2-2t_5)\partial_3\partial_6\). The third term is also a sum over all the edges \(i\) and the inner sum contains the two triangles that can be formed containing the edge \(i\); for \(i=1\) this means \((\partial_3+\partial_5)(\partial_2+\partial_6)\). The very last term comes from the three pairs of opposite edges and has two degenerations in the generating function with sign distributions depending on whether the degenerations are on opposite or adjacent edges. In the case considered in , all boundary terms could be simplified to Eisenstein series or products thereof by virtue of [\[eq:CabcSpec\]](#eq:CabcSpec){reference-type="eqref" reference="eq:CabcSpec"}; here the source terms are of a more complicated nature but still simpler compared to the full tetrahedral function. This can be seen in the examples above. As is shown in appendix [\[app:SL3C\]](#app:SL3C){reference-type="ref" reference="app:SL3C"}, the Laplace equation [\[eq:LG6\]](#eq:LG6){reference-type="eqref" reference="eq:LG6"} for the generating function can be rewritten using the quadratic Casimir of the split real \(\mf{sl}(3)\). Upon inserting the Casimir \[\begin{aligned}
\mathfrak{C}^2 &= \frac43 \mathfrak{D}^2 + 2 \mathfrak{D} + \sum_{V_{ijk}} (t_i^2+t_j^2+t_k^2-2t_i t_j-2 t_j t_k-2t_k t_i) (\partial_i \partial_j +\partial_j \partial_k +\partial_k\partial_i)\nonumber\\ & \quad\quad\quad + \sum_{i=1}^3 ((t_{i+1}-t_{i+2}-t_{i+4} +t_{i+5})^2-4 t_i t_{i+3}) \partial_i \partial_{i+3}\. \end{aligned}\] that is derived in appendix [\[app:SL3C\]](#app:SL3C){reference-type="ref" reference="app:SL3C"}, we obtain \[\begin{aligned}
\label{eq:LapCas} \left( \Delta-\mf{C}^2 +\frac13\left(\mathfrak{D}+3\right)\left(\mathfrak{D}+6\right)\right) {\mathcal{W}} = {\mathcal{R}}\,, \end{aligned}\] where we have suppressed all dependence on the variables \(\tau\) and \(t_i\) of the generating function \({\mathcal{W}}\) and the remainder \({\mathcal{R}}\). Solving the spectral problem means finding the spectrum of the operator \[\begin{aligned}
\label{eq:L2C2} \mf{L}^2 = \mf{C}^2-\frac13\left(\mathfrak{D}+3\right)\left(\mathfrak{D}+6\right)\. \end{aligned}\] We note that the occurrence of \(\mf{sl}(3)\) is slightly misleading, there is no actual \(\mf{sl}(3)\) symmetry of the spectrum; what we will be interested in is the number of \(\mf{S}_4\) singlets in representations of \(\mf{sl}(3)\). This situation is generalisation of the case discussed in appendix [\[app:2loop\]](#app:2loop){reference-type="ref" reference="app:2loop"} for the sunset graph underlying the functions \(C_{(s,t,p)}\).
## General spectral considerations
We now try to find a basis of \(\mf{C}^2\)-eigenfunctions of homogeneous polynomials in the six \(t_i\) that transform under \(\mathfrak{S}_4\) in such a way that the polynomials are invariant. The action of \(\mf{S}_4\) is induced from that of \(\mf{sl}(3)\) mentioned above and exhibited in appendix [\[app:SL3C\]](#app:SL3C){reference-type="ref" reference="app:SL3C"}. That is, we are looking for functions that satisfy \[\begin{aligned}
\mathfrak{D} \mathcal{W}_{w,p_1,p_2} & = (w-6) \mathcal{W}_{w,p_1,p_2}\,,\nonumber\\ \mathfrak{C}^2 \mathcal{W}_{w,p_1,p_2} & = \frac13 (p_1^2+p_1p_2+p_2^2+3p_1+3p_2)\mathcal{W}_{w,p_1,p_2} \equiv\lambda_{p_1,p_2} \mathcal{W}_{w,p_1,p_2} \end{aligned}\] and are \(\mf{S}_4\)-invariant. We here are using the \(\mf{sl}(3)\) quadratic Casimir operator \(\mf{C}^2\) that was defined in [\[eq:CasSL3\]](#eq:CasSL3){reference-type="eqref" reference="eq:CasSL3"} that is normalised such that when acting on an \(\mf{sl}(3)\) representation with Dynkin labels \([p_1,p_2]\) it has eigenvalue \(\lambda_{p_1,p_2} = \frac13 (p_1^2+p_1p_2+p_2^2+3p_1+3p_2)\). We note that the dimension of the \(\mf{sl}(3)\) representation with labels \([p_1,p_2]\) is given by \[\begin{aligned}
\dim\, [p_1,p_2] = \frac{(p_1+p_2+2)(p_1+1)(p_2+1)}{2}\. \end{aligned}\] The eigenvalue \(k\) of the scaling operator \(\mf{D}\) is related to the weight \(w=\sum_i s_i\) of \(C\!\!\raisebox{-0.3\height}{\six{s_1}{s_2}{s_3}{s_4}{s_6}{s_5}}\) by \[\begin{aligned}
k = w-6\. \end{aligned}\] The value \(k\) corresponds to the degree of the homogeneous polynomial in the \(t_i\). The operators \(\mf{D}\) and \(\mf{C}^2\) commute and the eigenvalue of the modular Laplacian is then \[\begin{aligned}
\label{eq:CD} \left( \Delta-\lambda_{p_1,p_2} +\frac13w(w-3)\right) \mathcal{W}_{w,p_1,p_2} = \mathcal{R}_{w,p_1,p_2} \end{aligned}\] according to [\[eq:L2C2\]](#eq:L2C2){reference-type="eqref" reference="eq:L2C2"}. We note that \(\mf{D}\) and \(\mf{C}^2\) do not form a complete set of commuting semi-simple operators. There are still degeneracies in the eigenspace labelled by \((w,p_1,p_2)\). The form of the right-hand side above depends on which particular eigenfunction in the \((w,p_1,p_2)\) we are considering. Mimicking the analysis of the two-loop sunset graph in appendix [\[app:2loop\]](#app:2loop){reference-type="ref" reference="app:2loop"}, we need to first identify the correct representations of \(\mf{sl}(3)\). The representation on six variables \(t_i\) is the irreducible six-dimensional representation that we choose as \([2,0]\) by some labelling convention for the nodes of the \(A_2\) Dynkin diagram. For homogeneous polynomials of degree \(k\) we need to form its symmetric tensor product series. We first compute the total number of \(\mf{S}_4\)-invariant functions for a given degree \(k=w-6\) of the polynomial. This can be done by applying Molien's theorem to the six-dimensional representation of \(\mf{S}_4\) and gives the following generating function \[\begin{aligned}
\frac{1-q+q^2+q^4+q^6-q^7+q^8}{(1-q)^6 (1+q)^2 \left(1+q^2\right) \left(1+q+q^2\right)^2} \end{aligned}\] that is also documented as series [A003082](https://oeis.org/A003082) in the OEIS . From this one can construct the total number of \(\mf{S}_4\) singlets at a given order \[\begin{aligned}
\label{eq:tetspec} \arraycolsep=2mm \begin{array}{c|cccccccc} k & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7\\\hline \mf{S}_4 \textrm{ singlets in \(S^k([2,0])\)} & 1 & 1 & 3 & 6 & 11 & 18 & 32 & 48 \end{array} \end{aligned}\] As \(w=k+6\), we recognise the single tetrahedral modular graph function at weight \(6\) and the single tetrahedral modular graph function at weight \(7\) discussed above. In order to separate the total number of eigenfunctions at weight \(w=k+6\) into the Casimir eigenspaces of the representation \([p_1,p_2]\) we need to consider the \(\mf{sl}(3)\) representations occurring in the \(k\)-th symmetric tensor power of the six-dimensional representation \([2,0]\) of the \(t_i\) variables. This is given by \[\begin{aligned}
\label{eq:Sk20} S^k ([2,0]) = \bigoplus_{i=0}^{\lfloor\frac{k}{3}\rfloor} \bigoplus_{j=0}^{\lfloor\frac{i}{2}\rfloor} [2k-6i,6j] \oplus \bigoplus_{i=0}^{\lfloor\frac{k-2}{3}\rfloor} \bigoplus_{j=0}^{\lfloor\frac{i}{2}\rfloor}[2k-6i-4,6j+2] \oplus \bigoplus_{i=0}^{\lfloor\frac{k-1}{3}\rfloor} \bigoplus_{j=0}^{\lfloor\frac{i-1}{2}\rfloor}[2k-6i-2,6j+4]. \end{aligned}\] All these representations of \(\mf{sl}(3)\) occur with multiplicity one. The only degeneracies in the Casimir eigenvalues arise for representations related by the outer automorphism of \(\mf{sl}(3)\), *i.e.*, only \([p_1,p_2]\) and \([p_2,p_1]\) have the same Casimir eigenvalue, otherwise all Casimir values are different. Unfortunately, compared to the two-loop case of appendix [\[app:2loop\]](#app:2loop){reference-type="ref" reference="app:2loop"}, we do not have a direct description of all \([p_1,p_2]\) as symmetric powers of some simple representation. A notable exception is again given by the symmetric powers of the fundamental (and anti-fundamental) representation: \[\begin{aligned}
S^p([1,0])= [p,0] \quad\quad \textrm{(\(S^p([0,1]) = [0,p]\))}\. \end{aligned}\] As a representation of \(\mf{S}_4\) the three-dimensional fundamental representation of \(\mf{sl}(3)\) is the standard representation and one can choose as generators for example the matrices \[\begin{aligned}
\begin{pmatrix} 0 & 0 & 1\\ 0 & 1 & 0\\-1 & 1 &0 \end{pmatrix} \,,\quad \begin{pmatrix} 0 & 0 &-1\\ 0 &-1 & 0\\-1 & 0 &0 \end{pmatrix} \,,\quad \begin{pmatrix} 0 & 0 & 1\\ 1 &-1 & 1\\ 1 & 0 &0 \end{pmatrix}\. \end{aligned}\] Molien's theorem gives the number of \(\mathfrak{S}_4\) invariants in such representations as being generated by \[\begin{aligned}
\frac{1-q^3+q^6}{(1-q)^3(1+q)^2(1+q^2)(1+q+q^2)}\. \end{aligned}\] For low \(p\) one has for the number of \(\mf{S}_4\) singlets in \([p,0]\) (or equivalently \([0,p]\)) \[\begin{aligned}
\arraycolsep=2mm \begin{array}{c|ccccccc} p & 0 & 1 & 2 & 3 & 4 & 5 & 6\\\hline \mf{S}_4 \textrm{ singlets in \([p,0]\)} & 1 & 0 & 1 & 0 & 2 & 0& 3 \end{array} \end{aligned}\] (The series starts to be non-zero for odd \(p\) soon after this.) At *even* \(p\) this series agrees with \[\begin{aligned}
\frac{1}{(1-q^2)^3(1+2q^2+2q^4+q^6)}= \frac{1}{(1-q^2)(1-q^4)(1-q^6)} \end{aligned}\] as can be checked using a double extension of \(\mathfrak{S}_4\). The even \(p\) values are the only ones of interest to us, so we might as well work with this simpler series that is identical to the two-loop case. For more general \([p_1,p_2]\) it is not quite clear how to get the right generating series. In table [1](#tab:S4sing){reference-type="ref" reference="tab:S4sing"}, we present the number of \(\mf{S}_4\) invariants for small even \(p_1\geq p_2\). This inequality is sufficient due to the outer automorphism of \(\mf{sl}(3)\) and only even values of the \(p_i\) can arise in [\[eq:Sk20\]](#eq:Sk20){reference-type="eqref" reference="eq:Sk20"}.
The eigenvalue of the modular Laplacian at a given weight \(w\) then has to be calculated using [\[eq:CD\]](#eq:CD){reference-type="eqref" reference="eq:CD"}.
## Eigenfunctions and Laplace equations at low weight
In this section, we give some more examples of eigenvalues and eigenfunctions of the modular Laplacian acting on tetrahedral modular graph functions. We stress that we are using the terms 'eigenvalues' and 'eigenfunctions' loosely as the corresponding Laplace equations are typically inhomogeneous but the right-hand side source is of lower complexity. The explicit eigenfunctions of the modular Laplacian at low weights can be constructed using a basis of \(\mf{S}_4\)-invariant homogeneous polynomials of degree \(k\). We list as examples the linear and quadratic invariant homogeneous polynomials: \[\begin{aligned}
k=1:&& p_1(t) &= t_1+t_2+t_3+t_4+t_5+t_6\,,\nonumber\\ k=2: && p_2^{(1)}(t)&=t_1^2+t_2^2+t_3^2+t_4^2+t_5^2+t_6^2\,,\nonumber\\ && p_2^{(2)}(t)&=t_1 t_2 + t_1 t_3 + t_2 t_3 + t_2 t_4 + t_3 t_4 + t_1 t_5 + t_2 t_5 + t_4 t_5 + t_1 t_6 + t_3 t_6 + t_4 t_6 + t_5 t_6\,,\nonumber\\ && p_2^{(3)}(t)&=t_1t_4+t_3t_5 + t_2t_6\. \end{aligned}\] Similar bases of \(\mf{S}_4\) invariant homogeneous polynomials can be generated at any degree easily. The procedure for finding explicit eigenfunctions of \(\mf{L}^2\) is then to first diagonalise the action of the \(\mf{sl}(3)\) Casimir \(\mf{C}^2\) on the polynomials and then convert this to inhomogeneous Laplace equations for combinations of tetrahedral modular graph functions.
### Weight \(7\)
The linear polynomial \(p_1\) is an eigenfunction of the \(\mf{sl}(3)\) Casimir \(\mf{C}^2\) given in [\[eq:CasSL3\]](#eq:CasSL3){reference-type="eqref" reference="eq:CasSL3"} according to \[\begin{aligned}
\left(\mf{C}^2-\frac{10}{3}\right) p_1 = 0\. \end{aligned}\] The corresponding tetrahedral modular graph function is \(C\raisebox{-0.5\height}{\six{2}{1}{1}{1}{1}{1}}\) with Laplace equation given already in [\[eq:L111111\]](#eq:L111111){reference-type="eqref" reference="eq:L111111"}.
### Weight \(8\)
For \(k=2\) the following are explicit eigenfunctions of the Casimir operator [\[eq:CasSL3\]](#eq:CasSL3){reference-type="eqref" reference="eq:CasSL3"} in the normalisation given there:
This is the first time a degeneracy arises in the spectrum and we have chosen some particular simple basis. These \(\mf{S}_4\) invariant eigenfunctions [\[eq:deg2ef\]](#eq:deg2ef){reference-type="eqref" reference="eq:deg2ef"} of the \(\mf{sl}(3)\) Casimir \(\mf{C}^2\) can be translated into combinations of tetrahedral modular graph functions as follows
As can be seen, all right-hand sides contain the function \(C\raisebox{-0.5\height}{\six{-1}{3}{3}{1}{1}{1}}\). This function cannot be reduced by means of the simplification rules given in [\[eq:simpT\]](#eq:simpT){reference-type="eqref" reference="eq:simpT"}. However, we expect there to be an additional simplification rule that we have not been able to derive and that would simplify this function.
### Weight \(9\)
For \(k=3\) one has in total six eigenfunctions according to [\[eq:tetspec\]](#eq:tetspec){reference-type="eqref" reference="eq:tetspec"}. These separate into the \(\mf{L}^2\) eigenvalues \(-18\), \(-10\) and \(0\) with degeneracies \(1\), \(2\) and \(3\), respectively. For \(\mf{L}^2=-18\) the eigenfunction is given by \[\begin{aligned}
\mf{L}^2=-18:\quad& C_{\six{2}{2}{1}{1}{2}{1}} +3C_{\six{2}{2}{1}{2}{1}{1}} \. \end{aligned}\] For \(\mf{L}^2=-10\) one has the following basis of two eigenfunctions \[\begin{aligned}
\mf{L}^2=-10:\quad& C_{\six{2}{2}{1}{1}{2}{1}} +2C_{\six{2}{2}{1}{2}{1}{1}} +2C_{\six{3}{1}{1}{2}{1}{1}} \,, \nonumber\\ &C_{\six{2}{2}{2}{1}{1}{1}} +C_{\six{2}{2}{1}{1}{2}{1}} +3C_{\six{2}{2}{1}{2}{1}{1}}+4C_{\six{3}{2}{1}{1}{1}{1}} +2C_{\six{3}{1}{1}{2}{1}{1}} \. \end{aligned}\] For \(\mf{L}^2=0\) one has the follow three independent eigenfunctions \[\begin{aligned}
\mf{L}^2=-10:\quad& C_{\six{2}{2}{2}{1}{1}{1}} +C_{\six{2}{2}{1}{2}{1}{1}} +4C_{\six{3}{2}{1}{1}{1}{1}}+5C_{\six{4}{1}{1}{1}{1}{1}}\,,\nonumber\\ &-C_{\six{2}{2}{1}{1}{2}{1}} +3C_{\six{2}{2}{1}{2}{1}{1}}+3C_{\six{3}{1}{1}{2}{1}{1}}\,,\nonumber\\ & 3C_{\six{2}{2}{2}{1}{1}{1}} -2C_{\six{2}{2}{1}{1}{2}{1}}-6C_{\six{2}{2}{1}{2}{1}{1}} -18C_{\six{3}{2}{1}{1}{1}{1}}-9C_{\six{3}{1}{1}{2}{1}{1}}\. \end{aligned}\] We do not spell out the right-hand sides of the inhomogeneous Laplace equations as they are rather involved but note that, similar to [\[eq:Inhdeg2\]](#eq:Inhdeg2){reference-type="eqref" reference="eq:Inhdeg2"} they can involve tetrahedral modular graph functions on the right-hand side with where one edge has value \(-1\). Such terms possibly simplify.
### Weights \(10\), \(11\) and \(12\)
For weights \(10\), \(11\) and \(12\) we only present table [2](#tab:higherwt){reference-type="ref" reference="tab:higherwt"} of the degeneracies of the eigenvalues of \(\mf{L}^2\) and do not list the explicit eigenfunctions as they become rather involved.
The numbers in table [2](#tab:higherwt){reference-type="ref" reference="tab:higherwt"} can also be derived from [\[eq:Sk20\]](#eq:Sk20){reference-type="eqref" reference="eq:Sk20"}, [\[eq:L2C2\]](#eq:L2C2){reference-type="eqref" reference="eq:L2C2"} and table [1](#tab:S4sing){reference-type="ref" reference="tab:S4sing"}. We have additionally determined the corresponding eigenfunctions and checked that their inhomogeneous Laplace equations contain only less complex modular graph functions or tetrahedral modular graph functions that have one edge with value \(-1\).
### More Laplace equations
In this section we present some additional Laplace equations, where the Laplacian is not diagonalised as in the previous examples but instead the combinations are chosen such that there are no functions with a value \(-1\) on any edge remaining on the right-hand side. These together with the previous ones could be useful for finding the integrated versions of the corresponding amplitudes. In general, there remain tetrahedral modular graph functions with a similar complexity on the right-hand side. For weight \(w=8\) one has \[\begin{aligned}
\Delta \biggr( C_{\six{2}{1}{1}{2}{1}{1}} + 4 C_{\six{2}{2}{1}{1}{1}{1}} \biggr) &=-12 C_{\six{2}{1}{1}{2}{1}{1}}-40 C_{\six{2}{2}{1}{1}{1}{1}} + 48 C_{\three{1}{2}{5}} + 136 C_{\three{1}{3}{4}} + 60 C_{\three{2}{2}{4}} \nonumber\\ &\quad+44 C_{\three{2}{3}{3}} -8 C_{\five{1}{2}{1}{2}{2}}-36 E^2_{4}-16 E_{3} E_{5} + 52 E_{8} \. \end{aligned}\] The Laplacian acting on the function \(C\!\raisebox{-0.4\height}{\six{3}{1}{1}{1}{1}{1}}\) does not produce any \(-1\) values either. For weight \(w=9\), there are four combinations that do not produce any \(-1\) edges after application of the Laplace operator. Besides the functions \(C\!\raisebox{-0.4\height}{\six{4}{1}{1}{1}{1}{1}}\) and \(C\!\raisebox{-0.4\height}{\six{2}{2}{2}{1}{1}{1}}\) they are \[\begin{aligned}
( \Delta + 18 ) \biggl( C_{\six{1}{1}{1}{2}{2}{2}} + 3 C_{\six{1}{1}{2}{1}{2}{2}} \biggr)= 72 C_{\three{1}{3}{5}} + 45 C_{\three{1}{4}{4}} + 36 C_{\three{2}{2}{5}} + 72 C_{\three{2}{3}{4}}-36 E_4 E_5 + 36 E_9 \end{aligned}\] and \[\begin{aligned}
\Delta \left(4 C_{\six{1}{1}{1}{1}{2}{3}}+C_{\six{1}{1}{2}{1}{3}{1}}\right) &= -4 C_{\six{1}{1}{2}{2}{1}{2}} +14 C_{\six{1}{1}{2}{1}{2}{2}}-16 C_{\six{1}{1}{1}{1}{2}{3}}+6 C_{\six{1}{1}{1}{2}{2}{2}}-16 C_{\five{3}{1}{3}{1}{1}}\nonumber\\ &\quad -16 C_{\five{3}{1}{2}{2}{1}}-16 C_{\five{2}{1}{2}{3}{1}}-4 C_{\five{2}{2}{2}{2}{1}}-4 C_{\five{2}{1}{1}{2}{3}}\\ &\quad-12 E_4 C_{\three{1}{2}{2}}+12 E_4 C_{\three{1}{2}{4}}+84 C_{\three{1}{2}{6}}+128 C_{\three{1}{3}{5}} +96 C_{\three{1}{4}{4}}\nonumber\\ &\quad +48 C_{\three{2}{2}{5}}+34 C_{\three{2}{3}{4}}+22 C_{\three{3}{3}{3}}-48 E_4 E_5-20 E_3 E_6+68 E_9\.\nonumber \end{aligned}\] For weight \(w=10\), there are seven combinations that do not produce any \(-1\). For weight \(w=11\) there are \(11\) and for weight \(w=12\) there are \(19\) such combinations. | {'timestamp': '2017-10-05T02:08:51', 'yymm': '1706', 'arxiv_id': '1706.01889', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01889'} |
null | null |
# Introduction
##
The setting of finite groups acting on categories is a well-studied ground, see e.g. and references therein. A useful way to define the action is to require for every \(g \in G\) an autoequivalence \(\rho_g\colon \mathcal C \to \mathcal C\) together with a choice of isomorphisms \(\rho_g \rho_h \simeq \rho_{gh}\) satisfying a cocycle condition, see [2.1](#def-action){reference-type="ref" reference="def-action"}. One would then study the category of equivariant objects \(\mathcal C^G\), see [2.4](#def-equiv){reference-type="ref" reference="def-equiv"}.
## {#ex-Db}
For instance, if \(\mathcal C = \mathcal D^b(X)\) is the derived category of coherent sheaves on a variety \(X\) then a \(G\)-action on \(X\) induces a \(G\)-action on \(\mathcal C\), and furthermore \(\mathcal C^G\) can be interpreted as the derived category of coherent sheaves on the quotient stack \(X/G\).
## {#section-1}
The main goal of this paper is to give a direct proof of the Theorem of Elagin stating that if \(\mathcal C = \langle \mathcal A, \mathcal B \rangle\) is a semi-orthogonal decomposition of triangulated categories and \(G\) is finite group acting on \(\mathcal C\) by triangulated autoequivalences in such a way that the category of equivariant objects \(\mathcal C^G\) is triangulated and preserving \(\mathcal A\) and \(\mathcal B\), then there is a semi-orthogonal decomposition \(\mathcal C^G = \langle \mathcal A^G, \mathcal B^G \rangle\), see Theorem [6.2](#thm-elagin){reference-type="ref" reference="thm-elagin"}. In the setup of [1.2](#ex-Db){reference-type="ref" reference="ex-Db"} this Theorem is often quite useful in constructing semiorthogonal decompositions for the quotient stack \(\mathcal D^b(X/G)\) from semiorthogonal decompositions of \(\mathcal D^b(X)\).
## {#section-2}
In our proof we construct the functors \(\mathcal C^G \to \mathcal A^G\) and \(\mathcal C^G \to \mathcal A^G\) adjoint to the inclusion functors. The key step in the proof is to show that if \(\Phi\colon \mathcal A \to \mathcal C\) is a \(G\)-equivariant functor which admits a left or right adjoint functor \(\Psi\), then \(\Psi\) is automatically equivariant: see Proposition [3.13](#adj-mates){reference-type="ref" reference="adj-mates"}.
## {#section-3}
We also prove that every \(G\)-action on a category \(\mathcal C\) is \(G\)-weakly equivalent to a strict \(G\)-action, that is to an action satisfying \(\rho_g \rho_h = \rho_{gh}\), see Theorem [5.4](#strict){reference-type="ref" reference="strict"}. This is analogous to the Coherence Theorem for monoidal categories: every monoidal category is equivalent to a strict monoidal category, see e.g..
## {#section-4}
In order to formulate and prove these facts we need to develop the language of \(G\)-functors, \(G\)-natural transformations and so on. Perhaps relevant definitions and constructions are well-known to experts but we include these for completeness as we could not find the reference that fits our purpose.
## {#section-5}
All categories, functors etc are \(k\)-linear where \(\mathrm{char}(k) = 0\). Groups acting on categories are finite and we denote by \(1 \in G\) the neutral element of the group. We use the symbol "\(\circ\)" to denote vertical composition of natural transformations of functors, the other types of compositions are denoted by concatenation.
# \(G\)-categories and equivariant objects
## {#def-action}
By a **\(G\)-action on \(\mathcal C\)** we mean the following data:
- For each element \(g \in G\) an autoequivalence \(\rho_g\colon \mathcal C \to \mathcal C\)
- For each pair \(g,h \in G\) an isomorphism of functors \[\phi_{g,h}\colon \rho_g \rho_h \cong \rho_{g h}.\]
The data must satisfy the following associativity axiom: for all \(g,h,k \in G\) the diagram of functors \(\mathcal C \to \mathcal C\) is commutative: \[\rho_g \rho_h \rho_k \ar[d]_{\phi_{g,h}\rho_k} \ar[r]_{\rho_g \phi_{h,k}} & \rho_g \rho_{h k} \ar[d]_{\phi_{g,hk}} \\ \rho_{gh} \rho_k \ar[r]_{\phi_{gh,k}} & \rho_{ghk} }\]
## {#section-6}
It follows from the definition that there is an isomorphism of functors \[\phi_1\colon \rho_1 \simeq id\] obtained by post-composing \[\phi_{1,1}\colon \rho_1 \rho_1 \to \rho_1\] with \(\rho_1^{-1}\). That is we have \[\phi_{1,1} = \rho_{1} \phi_{1}.\] Furthermore one can show that \(\phi_1\) satisfies: \[\begin{aligned} \phi_{g,1} &= \rho_g \phi_1\colon \rho_g \rho_1 \to \rho_g \\ \phi_{1,g} &= \phi_1 \rho_g\colon \rho_1 \rho_g \to \rho_g \\ & \\ \end{aligned}\] so that definition [2.1](#def-action){reference-type="ref" reference="def-action"} coincides with that of. On the other hand if one asks for \(\phi_1\) to be the identity transformation, one gets a slightly stronger definition of a \(G\)-descent datum of.
## {#section-7}
Using the language of monoidal functors one can give a very concise definition of a group acting on a category. For that consider \(G\) as a monoidal category: \(G\) is discrete as a category and its monoidal structure defined by \[g \otimes h = gh\] \[id_g \otimes id_h = id_{gh}.\] Now a \(G\)-action on \(\mathcal C\) amounts to the same thing as an action of monoidal category \(G\) on \(\mathcal C\), i.e. a weak monoidal functor \[\rho\colon G \to [\mathcal C, \mathcal C]\] where on the right is the category of functors \(\mathcal C \to \mathcal C\) with monoidal structure given by composing functors.
## {#def-equiv}
One defines the **category of \(G\) equivariant objects \(\mathcal C^G\)** as follows: objects of \(\mathcal C^G\) are linearized objects, i.e. objects \(c \in \mathcal C\) equipped with isomorphisms \[\theta_g\colon c \to \rho_g (c), \; g \in G\] satisfying the condition that the diagrams are commutative: \[c \ar[d]_{\theta_{gh}} \ar[r]_{\theta_g} & \rho_g(c) \ar[d]_{\rho_g \theta_h} \\ \rho_{gh}(c) & \ar[l]_{\phi_{g,h}(c)} \rho_g (\rho_h (c)) }\] Morphisms of equivariant objects consist of those morphisms of the underlying objects in \(\mathcal C\) which commute with all \(\theta_g\), \(g \in G\).
# \(G\)-functors and \(G\)-natural transformations
## {#pentagon}
Given two categories \(\mathcal C\), \(\mathcal D\) with \(G\)-actions and a functor \(\Phi\colon \mathcal C \to \mathcal D\), \(\Phi\) is called a **right lax \(G\)-functor** if there are given natural transformations \[\delta_g\colon \rho_g \Phi \to \Phi \rho_g\] such that the two natural transformations \(\rho_g \rho_h \Phi \to \Phi \rho_{gh}\) coincide: \[& & \rho_{gh} \Phi \ar[drr]^{\delta_{gh}} & & \\ \rho_g\rho_h\Phi \ar[urr]^{\phi_{g,h} \Phi} \ar[dr]_{\rho_g \delta_h} & & & & \Phi\rho_{gh} \\ & \rho_g\Phi\rho_h \ar[rr]_{\delta_g \rho_h} & & \Phi\rho_{g}\rho_h \ar[ur]_{\Phi \phi_{g,h}} & & \\ }\] This commutative diagram is called the pentagon axiom. Similarly \(\Phi\) is called a **left lax \(G\)-functor** if there are given natural transformations \[\delta_g\colon \Phi \rho_g \to \rho_g \Phi\] satisfying the dual pentagon axiom. A right (or left) lax \(G\)-functor \(\Phi\) is called a **weak \(G\)-functor** if all \(\delta_g\) are isomorphisms. The following lemma is a useful criterion for a weak \(G\)-functor.
## Lemma {#lem-weak}
Let \(\Phi\) be a right (or left) lax \(G\)-functor. The following conditions are equivalent:
1. The natural transformation \(\delta_1\colon \rho_1 \Phi \to \Phi \rho_1\) is an isomorphism.
2. \(\Phi\) satisfies the identity element axiom: \[\Phi \phi_1 \circ \delta_1 = \phi_1 \Phi\colon \rho_1 \Phi \to \Phi.\]
3. \(\Phi\) is a weak \(G\)-functor.
## Proof
Implications \((3) \implies (1)\), \((2) \implies (1)\) are obvious. Let us prove that \((1) \implies (3)\). Consider the case of the right lax \(G\)-functor. Applying the pentagon axiom to the pair \((g^{-1},g)\) gives: \[\delta_{g^{-1}} \rho_g \circ \rho_{g^{-1}} \delta_g = \Phi \phi_{g^{-1},g}^{-1} \circ \delta_1 \circ \phi_{g^{-1},g} \Phi.\] Since the natural transformation on the right-hand side is an isomorphism (note that \(\delta_1\) is an isomorphism by the identity element axiom) and \(\rho_{g}\), \(\rho_{g^{-1}}\) are equivalences, it follows that \(\delta_{g^{-1}}\) is left invertible and \(\delta_g\) is right invertible. Thus we see that all \(\delta_g\) are isomorphisms. Now we prove \((1) \implies (2)\). Consider the natural transformation \[\varepsilon = \Phi \phi_1 \circ \delta_1 \circ \phi_1^{-1} \Phi\colon \Phi \to \Phi.\] We are given that \(\varepsilon\) is an isomorphism and we need to prove that \(\varepsilon\) is in fact an identity. We use Lemma [3.4](#G-compos){reference-type="ref" reference="G-compos"} applied to the trivial group \(H := \{1\}\) and the composition \[(\mathcal C, id) \underset{(id, \phi_1)}{\to} (\mathcal C, \rho_1) \underset{(\Phi, \delta_1)}\to (\mathcal D, \rho_1) \underset{(id, \phi_1^{-1})}\to (\mathcal D, id)\] which gives a lax \(G\)-functor \[(\mathcal C, id) \underset{(\Phi, \varepsilon)}{\to} (\mathcal D, id).\] The pentagon axiom for this functor yields \[\varepsilon^2 = \varepsilon\] and we deduce that \(\varepsilon = id\).
## Lemma {#G-compos}
If \((\Phi, \delta^\Phi)\colon \mathcal C \to \mathcal D\), \((\Psi, \delta^\Psi)\colon \mathcal D \to \mathcal E\) are right/left/weak \(G\)-functors, then their composition \((\Psi \Phi, \Phi \delta^{\Psi} \circ \delta^\Phi \Psi)\) is a right/left/weak \(G\)-functor. For the proof one needs to check that the composition satisfies the pentagon and/or the identity element axioms; this is a straightforward check.
## Lemma {#lemma}
A weak \(G\)-functor \(\Phi\colon \mathcal C \to \mathcal D\) induces a functor on the categories of equivariant objects \[\Phi^G\colon \mathcal C^G \to \mathcal D^G\] such that the following diagram is commutative \[\mathcal C^G \ar[d] \ar[r]^{\Phi^G} & \mathcal D^G \ar[d] \\ \mathcal C \ar[r]^{\Phi} & \mathcal D }\]
## Proof
For \((c, \theta) \in \mathcal C^G\) we define linearization on \(\Phi(c)\) as a composition of isomorphisms \[\Phi(c) \to \Phi \rho_g(c) \to \rho_g \Phi(c)\] of \(\Phi \theta_g\) with \(\delta_g\). It is now a standard check that \(\Phi(c)\) becomes an equivariant object and that \(\Phi^G\) is a functor.
## Definition
A natural transformation between two weak \(G\)-functors \(\mu\colon \Phi_1 \to \Phi_2\colon \mathcal C \to \mathcal D\) is called a \(G\)-natural transformation if for every \(g \in G\) the following diagram commutes: \[\rho_g \Phi_1 \ar[d]_{\delta_{1,g}} \ar[r]^{\rho_g \mu} & \rho_g \Phi_2 \ar[d]_{\delta_{2,g}} \\ \Phi_1 \rho_g \ar[r]^{\mu \rho_g} & \Phi_2 \rho_g. }\]
## Lemma {#lem-Gnat}
A \(G\)-natural transformation \(\mu\) between two weak \(G\)-functors \(\Phi_1, \Phi_2\colon \mathcal C \to \mathcal D\) induces a natural transformation \(\mu^G\colon \Phi_1^G \to \Phi_2^G\).
## Proof
To prove that \(\mu\) descends to a natural transformation \(\mu^G\colon \Phi_1^G \to \Phi_2^G\) we check that for every \((c, \theta) \in \mathcal C^G\) the morphism \(\mu\colon \Phi_1(c) \to \Phi_2(c)\) commutes with linearizations: \[\rho_g \Phi_1(c) \ar[d]_{\delta_1}^{\simeq} \ar[r]^{\rho_g \mu(c)} & \rho_g \Phi_2(c) \ar[d]_{\delta_2}^{\simeq} \\ \Phi_1 \rho_g(c) \ar[r]^{\mu \rho_g(c)} & \Phi_2 \rho_g(c) \\ \Phi_1 (c) \ar[u]_{\Phi_1 \theta_g} \ar[r]^{\mu(c)} & \Phi_2(c) \ar[u]_{\Phi_2 \theta_g}\\ }\] The transformation \(\mu^G\) is natural since the original transformation \(\mu\) is natural and the forgetful functor \(\mathcal C^G \to \mathcal C\) is faithful.
## Definition
Two weak \(G\)-functors \(\Phi\colon \mathcal C \to \mathcal D\), \(\Psi\colon \mathcal D \to \mathcal C\) are called **\(\mathbf{G}\)-adjoint** if they are adjoint and the unit \(\varepsilon\colon id \to \Phi \Psi\) and counit \(\eta\colon \Psi \Phi \to id\) of the adjunction are \(G\)-natural transformations.
## Lemma {#adj-descent}
A \(G\)-adjoint pair of functors \(\Phi\), \(\Psi\) induces an adjoint pair \(\Phi^G\), \(\Psi^G\) between the categories of equivariant objects.
## Proof
From [3.8](#lem-Gnat){reference-type="ref" reference="lem-Gnat"} it follows that we have natural transformations \(\varepsilon^G\colon id \to \Phi^G \Psi^G\), \(\eta^G\colon \Phi^G \Psi^G \to id\). The condition for \(\Psi\) and \(\Phi\) to be adjoint is that two compositions \[\Phi \eta \circ \varepsilon \Phi\colon \Phi \to \Phi \Psi \Phi \to \Phi\] and \[\eta \Phi \circ \Psi \varepsilon\colon \Psi \to \Psi \Phi \Psi \to \Psi\] are identities. Since the forgetful functor \(\mathcal C^G \to \mathcal C\) is faithful, the same holds for \(\Phi^G\), \(\Psi^G\).
## Proposition {#adj-mates}
A left or right adjoint \(\Psi\) to a weak \(G\)-functor \(\Phi\) can be made into a weak \(G\)-functor in such a way that \(\Psi\) and \(\Phi\) become \(G\)-adjoint.
## Proof
Let \(\Psi\) be the left adjoint to \(\Phi\colon \mathcal C \to \mathcal D\). We construct the structure of a left lax \(G\)-functor on \(\Psi\) using the structure of a right lax \(G\)-functor on \(\Phi\). Let \(\varepsilon\colon id \to \Phi \Psi\) and \(\eta\colon \Psi \Phi \to id\) be the unit and the counit of the adjunction. Given a right lax \(G\)-structure \(\delta_g\colon \rho_g \Phi \to \Phi \rho_g\) on \(\Phi\) we define the left lax \(G\)-structure \(\delta_g'\colon \Psi \rho_g \to \rho_g \Psi\) on \(\Psi\) as a mate of \(\delta_g\) with respect to the adjunction, , i.e. \[\delta_g' = \eta \rho_g \Psi \circ \Psi \delta_g \Psi \circ \Psi \rho_g \varepsilon\colon \Psi \rho_g \to \Psi \rho_g \Phi \Psi \to \Psi \Phi \rho_g \Psi \to \rho_g \Psi.\] The pentagon axiom can be expressed as an equality of certain compositions in the double category of, hence is preserved under taking mates by. Checking the identity axiom for \(\delta_1'\) is straightforward. Now by [3.2](#lem-weak){reference-type="ref" reference="lem-weak"} \(\Psi\) becomes a weak \(G\)-functor. The proof for right adjoints is analogous. We now need to prove that the unit and counit transformations \(\varepsilon\), \(\eta\) are \(G\)-natural. We do the proof for the unit \(\varepsilon\). We need to check that the following diagram commutes: \[\rho_g id \ar[d]_{=} \ar[r]^{\varepsilon} & \rho_g \Phi \Psi \ar[d]^{\delta_{\Phi\Psi, g}} \\ id \, \rho_g \ar[r]^{\varepsilon} & \Phi \Psi \rho_g. \\ }\] Here \(\delta_{\Phi\Psi}\) is defined using [3.4](#G-compos){reference-type="ref" reference="G-compos"}. Unraveling the definitions we are left with checking the diagram (where we use simplified notation for the natural transformations to denote the obvious compositions) \[\rho_g \ar[r]^{\varepsilon} \ar[d]_{\varepsilon} & \rho_g \Phi\Psi \ar[r]^{\delta_g}_{\simeq} \ar[d]_\varepsilon & \Phi \rho_g \Psi \ar[r]^{=} \ar[d]_{\varepsilon} & \Phi \rho_g \Psi \\ \Phi \Psi \rho_g \ar[r]^{\varepsilon} & \Phi\Psi\rho_g \Phi\Psi \ar[r]^{\delta_g}_{\simeq} & \Phi \Psi \Phi \rho_g \Psi \ar[ur]_\eta\\ }\] which is easily seen to commute.
## Corollary
Let \(\Phi\colon \mathcal C \to \mathcal D\) be a weak \(G\)-functor. Then the following conditions are equivalent:
1. \(\Phi\) is an equivalence of categories
2. There exists a weak \(G\)-functor \(\Psi\colon \mathcal D \to \mathcal C\) and \(G\)-natural isomorphisms \(\Psi \circ \Phi \simeq {\rm id}_\mathcal C\), \(\Phi \circ \Psi \simeq {\rm id}_\mathcal D\).
In this case we will call \(\Phi\) a **weak \(G\)-equivalence**.
## Proof
We only need to prove \((a) \implies (b)\) as the opposite implication is trivial. Let \(\Psi\colon \mathcal D \to \mathcal C\) be the quasi-inverse functor to \(\Phi\). In particular \(\Psi\) and \(\Phi\) are adjoint (both ways) so that by [3.13](#adj-mates){reference-type="ref" reference="adj-mates"} \(\Psi\) has a structure of a weak \(G\)-functor with compositions \(G\)-isomorphic to identity functors.
# Example: \(G\)-actions on the category of vector spaces
## {#section-8}
In this section we review a well-known example of how equivalence classes of \(G\)-actions on the category of \(k\)-vector spaces correspond bijectively to cohomology classes \(H^2(G, k^*)\).
## {#section-9}
Let \(\mathcal C = {\rm Vect_k}\) be the category of \(k\)-vector spaces, and let \(\rho\) be the \(G\)-action on \({\rm Vect_k}\). As every autoequivalence of \(\mathcal C\) is isomorphic to the identity functor, let us assume \(\rho_g = {\rm id}\) for every \(g \in G\). In this setup the data of the \(G\)-action \(\rho\) defined in [2.1](#def-action){reference-type="ref" reference="def-action"} is equivalent to specifying a cocycle \(\phi \in Z^2(G,k^*)\).
## {#section-10}
Consider two \(G\)-actions on \({\rm Vect_k}\) given by cocycles \(\phi, \phi' \in Z^2(G,k^*)\). For the \(G\)-actions to be equivalent there needs to exist a weak \(G\)-functor \[\Phi\colon ({\rm Vect_k}, \phi) \to ({\rm Vect_k}, \phi')\] which is an equivalence of categories. Then the pentagon axiom [3.1](#pentagon){reference-type="ref" reference="pentagon"} requires existence of an element \(\delta = (\delta_g)_{g \in G} \in Z^1(G,k^*)\) such that \(\phi'_{g,h} = \delta_g \delta_h \delta_{gh}^{-1} \phi_{g,h}\) for all \(g,h\). Thus \(G\)-categories \(({\rm Vect_k}, \phi)\) and \(({\rm Vect_k}, \phi')\) are equivalent if and only if \([\phi] = [\phi'] \in H^2(G, k^*)\).
## {#section-11}
The category of equivariant objects \(({\rm Vect_k}, \phi)^G\) is the category of \(\phi\)-twisted \(G\)-representations with objects given by vector spaces \(V\) together with isomorphism \(\theta_g\colon V \to V\) satisfying \(\theta_{gh} = \phi(g,h) \theta_{g} \theta_{h}\) and \(G\)-equivariant morphisms. In particular, if \(\phi\) is the trivial cocycle, so that \(G\)-action on \({\rm Vect_k}\) is trivial, \({\rm Vect_k}^G\) is the category of \(G\)-representations.
# Strictifying \(G\)-actions
## {#section-12}
Let \(\Omega(G)\) denote the category with one object for every element \(g \in G\) with \({\rm Hom}(g,g) = k\) and \({\rm Hom}(g,h) = 0\) for \(g \ne h\).
## {#section-13}
Let \(\mathcal C\) be a category with a \(G\)-action. Consider the category of weak \(G\)-functors and \(G\)-natural transformations from \(\Omega(G)\) to \(\mathcal C\) \[\mathcal C' = {\rm Hom}_{G}(\Omega(G), \mathcal C).\] We endow \(\mathcal C'\) with the strict \(G\)-action induced by the \(G\)-action on \(\Omega(G)\).
## {#section-14}
Explicitly the objects of \(\mathcal C'\) consist of families \((c_g \in \mathcal C)_{g \in G}\) together with isomorphisms \(\delta_{h,g}\colon \rho_h c_g \simeq c_{hg}\) satisfying the cocycle condition that two ways of getting an isomorphism \(\rho_k \rho_h c_g \simeq c_{khg}\) coincide. The morphisms from \((c_g)_{g \in G}\) to \((d_g)_{g \in G}\) are morphisms \(f_g\colon c_g \to d_g\) satisfying the condition that the two natural ways of forming a morphism \(\rho_h c_g \to d_{hg}\) coincide.
## Theorem {#strict}
The functor \(\Phi\colon \mathcal C' \to \mathcal C\) sending \((c_g)_{g \in G}\) to \(c_1\) is a weak \(G\)-equivalence. Hence, every \(G\)-action is weakly equivalent to a strict \(G\)-action.
## Proof
We need to check that \(\Phi\) has a structure of a weak \(G\)-functor and that \(\Phi\) is fully faithful and essentially surjective. The structure of a weak \(G\)-functor on \(\Phi\) is in fact simply given by the structure maps \(\delta_{h,g}\). That is we have functorial isomorphisms \[\rho_g \Phi(c) = \rho_g(c_1) \overset{\delta_{g,1}}{\to} c_g = \Phi \rho_g(c)\] and the pentagon axiom follows from the cocycle condition on \(\delta\). To check that \(\Phi\) is essentially surjective, one checks that for any \(c \in \mathcal C\) the family \((\rho_g(c))\) has a structure of an object from \(\mathcal C\). Furthermore, one can see that any object \((c_g)_{g \in G}\) is isomorphic to \((\rho_g(c_1))_{g \in G}\). Thus to check that \(\Phi\) is fully faithful, we may take two objects \((c_g)_{g \in G} = (\rho_g(c_1))\) and \((d_g)_{g \in G} = (\rho_g(d_1))\) and a morphism \(f_g\colon c_g \to d_g\) between them. It is then easy to see that \(f_g = \rho(f_1)\) and that conversely for any \(f_1\colon c_1 \to d_1\), the collection \(\rho_g(f_1)\) defines a morphism between \(c\) and \(d\).
# Elagin's Theorem
## {#section-15}
If \(\mathcal C\) is a triangulated category and \(G\) acts by triangulated autoequivalences, then \(\mathcal C^G\) is endowed with a shift functor and a set of distinguished triangles: these are the triangles that are distinguished after applying the forgetful functor \(\mathcal C^G \to \mathcal C\). Furthermore under some mild technical assumptions this gives \(\mathcal C^G\) the structure of a triangulated category, for instance existence of a dg-enhancement of \(\mathcal C\) is a sufficient condition for \(\mathcal C^G\) to be triangulated.
## Theorem {#thm-elagin}
Let \(\mathcal C = \langle \mathcal A, \mathcal B \rangle\) be a semi-orthogonal decomposition of triangulated categories. Let \(G\) act on \(\mathcal C\) by triangulated autoequivalences which preserve \(\mathcal A\) and \(\mathcal B\). Assume that the equivariant category \(\mathcal C^G\) is triangulated with respect to triangles coming from \(\mathcal C\). Then \(\mathcal A^G\), \(\mathcal B^G \subset \mathcal C^G\) are triangulated and there is a semi-orthogonal decomposition \[\mathcal C^G = \langle \mathcal A^G, \mathcal B^G \rangle.\]
## Proof
The existence of an adjoint pair between \(\mathcal C\) and \(\mathcal C^G\) implies that \(\mathcal B^G = {}^\perp \mathcal A^G\) and \(\mathcal A^G = {\mathcal B^G}^\perp\). In particular \(\mathcal A^G\) and \(\mathcal B^G\) are triangulated subcategories of \(\mathcal C^G\). Now in order to establish the semi-orthogonal decomposition \(\mathcal C^G = \langle \mathcal A^G, \mathcal B^G \rangle\) it suffices to show that the embedding \(i^G\colon \mathcal A^G \to \mathcal C^G\) has a left adjoint. This holds true by [3.13](#adj-mates){reference-type="ref" reference="adj-mates"}, [3.11](#adj-descent){reference-type="ref" reference="adj-descent"}: the functor \(i\colon \mathcal A \to \mathcal C\) is (strictly) \(G\)-equivariant, hence its left adjoint \(p\colon \mathcal C \to \mathcal A\) induces an adjoint \(p^G\) to the embedding \(i^G\colon \mathcal A^G \to \mathcal C^G\). | {'timestamp': '2017-06-07T02:05:37', 'yymm': '1706', 'arxiv_id': '1706.01714', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01714'} |
# Introduction {#sec1}
Last few years have witnessed a massive growth in consumer interest in smart devices \[1\]. These include smartphones, smart TVs, and recently smartwatches. Because of their increased penetration, smartwatches have claimed a handsome market share. A typical smartwatch contains a heart rate monitor, GPS, thermometer, camera and accelerometer. Thus, it can provide a variety of services. These services include temperature and pulse measurements and the number of calories consumed when performing a physical activity, such as walking, running, cycling and so forth \[2\]. Recently, smartwatches have emerged in our daily lives and like their smartphone counterparts, smartwatches do contain a gyroscope, magnetometer and accelerometer sensors to support similar capabilities and applications. These applications include health applications i.e. calories consumption that require sensor based physical activity recognition. A number of calories consumed during physical activities vary from one activity to another. Therefore, the first thing to do when estimating the number of calories consumed, is to first recognize which activity have been performed \[3\]. This work explores the idea of physical activity recognition on smartwatches using the embedded gyroscope, magnetometer and accelerometer sensors. The gyroscope, magnetometer and accelerometer sensors are ideal for physical activity recognition. We further discussed the detailed analysis of various feature banks and classification methods to seek the best settings for optimum performance on two different types of models such as impersonal and personal model over the smartwatch. Personal models referred as the classifier is built using the data only form one specific user whereas the impersonal model referred as the classifier is built using the data form every user except the one under study. To summarize, the main contributions of this work is towards answering the following questions:
1. Existing smartphone-based physical activity recognition studies have a number of features that were utilized for smartphone-based physical activity recognition; can they be used for smartwatch-based physical activity recognition as well?
2. Which of the most commonly used classifiers from the reference field provides the best recognition accuracy?
3. Does features permutation help to improve the accuracy?
4. Does features normalization help to improve the performance of the classifier?
5. Which model (Personal or Impersonal) works better for smartwatch-based physical activity recognition?
6. Does changing the window size affect the recognition accuracy?
Physical activity recognition in general using motion sensors especially using smartphones has been studied in recent years \[4-6\] and currently being studied extensively \[7-8\] but there are a few studies on physical activity recognition using smartwatches \[9-18\] in which the authors studied the role of the smartwatch in physical activity recognition system for health recommendation purposes. For instance, Guiry et al. \[9\] proposed physical activity recognition system to recognize nine different activities using five different classifiers. These activities include walking, standing, cycling, running, stair ascent, stair descent, elevator ascent and descent. However, the authors studied smartphone and smartwatch based activity recognition separately and did not combine the data from both devices. Furthermore, for real-time data collection, the author's used magnetometer, accelerometer and gyroscope sensors for smartphone and accelerometer for a smartwatch. Trost et al. \[10\] used sensors on hip and wrist to detect physical activities using logistic regression based classification method. In \[10\], Trost et al. just focused on showing the potential of using the wrist position for physical activity recognition for health recommendation system. Chernbumroong et al. \[11\] used a single wrist-worn accelerometer to detect five different physical activities for recommendation system. These activities include standing, sitting, lying, running and walking. However, Da-Silva et al. \[12\] used the same sensor to detect the eight different physical activities for recommendation system. In addition to the activities identified in \[11\] Da-Silva et al. include the activity of working on a computer too. The works \[13\] and \[14\] detect the eating activity using a Hidden Markov Model (HMM) with a wrist-worn accelerometer and gyroscope sensors. Ramos-Garcia identifies the eating activity by dividing it into further sub-activities in \[13\] which includes resting, drinking, eating, using utensil and others. Meanwhile, Dong et al. \[14\] proposed to differentiate eating periods from non-eating periods to identify the eating and not-eating activity. Sen et al. \[15\] used both accelerometer and gyroscope sensor data to recognize the eating and non-eating activities. Furthermore, the authors used smartwatch camera to take images of the food being consumed to analyze what a user is eating. School et al. \[16\] presented a feasibility study to detect and identify the smoking activity using accelerometer sensor where School et al. only reported a user-specific accuracy for smoking activity. Whereas, A. Parade et al. \[17\] used both accelerometer, magnetometer and gyroscope sensors to recognize smoking puffs where they just try to distinguished smoking activity from other activities. A similar work has been reported in \[18\] where the the authors used support vector machine (SVM) to recognize six different physical activities. These activities include standing, walking, writing, jacks, smoking and jogging. Kim et al. \[19\] proposed the collaborative classification approach for recognizing daily activities using smartwatches. The authors exploit a single off-the-shelf smartwatch to recognize five daily activities; namely eating, watching TV, sleeping, vacuuming and showering. For fair comparisons Kim et al. suggested using single and multiple sensors-based approaches to compare with their proposed approach with the overall accuracy of 91.5% with the improved recall rate up to 21.5% for each activity. Weiss et al. showed that accelerometers and gyroscopes sensors that sense user's movements and can help identify the user activity \[20\]. Weiss et al. propose to compare the smartphone and smartwatch-based user activity recognition in which smartwatch have the ability to recognize hand-based activities namely eating which is not possible through a smartphone. They showed that the smartwatch is able to recognize the drinking activity with 93.3% accuracy while smartphones achieve only 77.3%. Nurwanto et al. \[21\] proposed a daily life light sport exercise activities. These activities include squat jump, push up, sit up and walking. Nurwanto et al. focuses only on accelerometer-based signal captured by left-handed persons and the obtained signals were processed through sliding window approach with the k-nearest neighbor method and dynamic time warping as the main classifier. Furthermore, Nurwanto et al. proved that k-NN together with dynamic time warping method is an efficient method for daily exercise recognition with the accuracy of 76.67% for push up 96.69% for squat jump and 80% for sit up activity. They intentionally exclude the walking activity for their experimental process because of the random patterns. Al-Naffakh et al. \[22\] proposed the usefulness of using smartwatch motion sensor namely gyroscope and accelerometer to perform user activity recognition. Al-Naffakh et al. proved the natures of the signals captured are sufficiently discriminative to be useful in performing activity recognition. Different activities can be recognized using motion sensors at the wrist position i.e. drinking, eating, smoking, walking, walking down stairs, walking up stairs, running, jogging, elevator up, elevator down, doing regular exercises, writing etc. However there is no study that discusses these activities all together with the recommendations of optimum system settings. Therefore, to extend the existing works, we evaluate the combination of different feature banks and classification techniques for the reliable recognition of various activities. We also evaluate three different built-in motion sensors. Moreover, unlike the current trends, we also evaluate the effect of window size on each activity performed by the individual user and its impact on the recognition performance of simple and complex activities. Although the effect of window size on physical activity recognition has been studied by Huynh et al. in \[23\] but the windowing effect on complex activities is yet not fully explored. Moreover, unlike the recent works, we explore the use of both *Personal* and *Impersonal* models individually for each activity.
1. Personal model means the classifier is trained using the data only from one specific individual whereas, the Impersonal model means the classifier is built using the data from every user except the one under study.
2. For both models, one subject cross-validation process is adopted.
3. Finally, our study suggests the optimum system settings for any activity recognition system based on smartwatches.
This work analyzes the effects of combining gyroscope, magnetometer and accelerometer sensors data for physical activity recognition for different size of windows varying from one second to 12 seconds long window. We further analyze the effect of various sampling rates on the recognition performance in different scenarios. However, in our current study, we still have room for further improvement. For example, our current dataset is not balanced (as shown in Fig. 1) for all performed activities which might lead to the biased results towards majority classes. To handle the said problem to some extent, we used one subject cross-validation process for different models (personal and impersonal) which somehow overcome the class imbalance issue. This study also overcomes the limitations of overlapping while segmenting the raw data which forcefully improve the recognition performance because this overlap cause using some part of data in both training and testing which limiting the use of such techniques in a real-time environment. Moreover, we proposed and suggest some methods which helps to improve the recognition performance for those activities that have low recognition performance. The rest of the paper has the following organization. Section 2 describes the material and methods. Section 3 contains experimental settings. Section 4 presents the experimental results and discussion and finally section 5 concludes the paper with possible future directions.
# Materials and Methods {#sec2}
Today's smartwatches are equipped with a variety of motion sensors that are useful for monitoring device movements like tilt, rotate and shake. Some of these sensors are the ambient light sensor, accelerometer, compass, gyroscope and GPS sensors. For experimental study we only collected the raw signals using three different sensors such as accelerometer, magnetometer and gyroscope sensor.
## Data Collection {#sec3}
We collected the raw signals for five different activities from six users (three female and three male) having a mean age of twenty-five years old. The criterion for selecting the subjects is based on the gender because different genders exhibit different patterns when performing the same activity. These activities include walking, walking upstairs, walking downstairs, running and jogging. All subjects performed these activities twice on each day for more than a month. Therefore we collected raw data from the same user for same activity but performed on different days. The participants enrolled in our study are approved by the laboratory head because this is a formal prerequisite because our experiments involved human subjects and there is a negligible risk of injury. The involved subjects then asked to answer few nontechnical questions like gender, age, height, weight, left or right handed etc, which we used to characterize our study. Then the subjects were asked to fastens the smartwatch on their wrist and places a Bluetooth paired smartphone in their pocket. Both devices run a simple custom designed application that controls the data collection process and instructs the participant to first add their name and then select the activity from the list of five different activities and the sensor from three different sensors. Once the initial instructions were done, turn the smartphone screen off and place the smartphone into the pants pocket. The smartphone instructs the smartwatch running our paired data collection application to collect the raw signal at the 20Hz rate. Each of these sensors generates 3-dimensional signals and appends a time-stamp to the values. Every after five minutes the smartwatch sends the data to the smartphone and after a successful transmission, the smartphone vibrate to notify the user that the data collection process is successfully completed and they can stop the current activity.
## Feature Extraction {#sec4}
The embedded sensor generates three-dimensional time series signals which are highly fluctuating and oscillatory in nature. The oscillation and fluctuations make the physical activity recognition more difficult than other applications in nature. These raw signals are shown in Fig. 1 for individual activity performed by the single user. Therefore, it is compulsory to gather the nontrivial signals from raw data through a feature extraction process. The raw signals are divided into several equal sized windows to control the flow rate and pass fewer data to the system to extract the meaningful information. In \[24\] Khan et al. have presented a detailed analysis of using different kinds of features with a different number of samples per windows for smartphone-based physical activity recognition. From these raw signals, the extracted features are divided into two different feature banks. First feature bank include the features obtained by taking; average, median, variance, standard deviation, interquartile, autocorrelation, partial autocorrelation, coefficients of autoregressive model, coefficients of moving average model, coefficients of autoregressive moving average model and wavelet coefficients. Second feature bank includes the features obtained by taking; average acceleration, average absolute difference, standard deviation, average resultant acceleration and the average difference between peaks. We also calculate the binned distribution in which we determine what fraction of reading fall in a 10 equal-sized bins. This function generates 10 features. Every feature except average resultant acceleration is extracted for each axis. These features are extracted from each axis of the three-dimensional acceleration signal and in total for each window of acceleration data 43 (first-feature-bank) and 70 (second-feature-bank) features are extracted respectively. Prior to the feature extraction process moving average filter of order three is used for noise reduction.
## Classifiers {#sec5}
Khan et al. and Saputri et al. showed that support vector machines (SVM) and artificial neural networks (ANN) are superior to other traditional classification methods for user-independent physical activity recognition system in \[24-26\]. K-nearest neighbor (KNN) is also one of the most popular classification method used for smartphone-based physical activity recognition and user identification \[27\]. ELM \[28\], Naive Bayes \[29\], Adaptive learning/Ensemble learning (BAG) \[30\] and Decision tree \[31\] classifiers are also famous for smartphone-based physical activity recognition system. All these classification methods have shown several advantages in several fields most probably in their respective domains. Once again, given that ours is an exploratory study and the fact that each of these classifiers has shown a good performance in their respective studies. Instead of choosing one we decided to explore the use of all of these classifiers except neural networks in this work.
# Experimental Settings {#sec6}
The following experimental studies are performed using all the data to create both personal and impersonal models. In our first study, using a fixed size of 75 samples per window the performance of five classifiers are compared to the following three cases on both feature banks and for both personal and impersonal models:
- Randomly permuted features with normalization.
- Without random permutation and normalization.
- Randomly permuted features without normalization.
In the second study, the best setting for each classifier is chosen and tested multiple times while changing the number of samples in a window *i.e. {25, 50, 75, 100, 125, 150, 175, 200, 225, 250, 275, 300}* samples. The goal is to measure the effect of changing the window size on the performance of each classifier for both personal and impersonal model respectively. In our both studies, the training and testing data is randomly divided and classification results are obtained using 10-fold cross-validation process. The values used for different parameters of classifiers are as follows where all these parameters are carefully tuned and optimized.
- SVM with quadratic kernel function is used.
- KNN with Euclidean distance is used and K is set to 10.
- Adaptive Learning: Ensemble model (BAG) with tree model and the number of weak classifiers are set to 50.
- Decision Tree: The numbers of trees are set to 85.
# Experimental Results and Discussion {#sec7}
This section contains the obtained results for smartwatch-based physical activity recognition system. To validate our system we conduct different experiments. Our first experiment is for both personal and impersonal model-based physical activity recognition analyses. In this experiment, we conduct a detailed comparison of two different feature banks and five classifiers for both personal and impersonal models. Our second experiment explains the physical activity recognition process behavior with a different number of samples per window within the best settings obtained in our first experiment. Finally, the third experiment presents the single subject cross-validation recognition for each activity for both personal and impersonal models individually. All these experiments are based on 10-fold cross-validation process and all these experiments are carried out using Matlab R2014b installed on core i5 and 8GB of RAM machine. In the introduction section; we stated some important research questions which our explanatory study seeks to provide the basis of answers. The answers to those questions based on the obtained results are as follows: Based on our findings, yes the features used in existing smartphone-based physical activity recognition studies can also be used for the smartwatch-based physical activity recognition. If we consider normalized and random permutation feature space, classifiers perform much better opposed to the case of without normalization and random permutation in several cases. Table I-IV present the 98% confidence intervals about overall physical activity recognition rate. Using pairwise T-tests between groups of normalized/randomly permuted and unnormalized/non-permuted data at the 98% confidence level.
Statistically significant results are in boldface, shows decision trees and SVM classifiers are statistically better when normalized and randomly permuted features are used. This increase was as much as almost 50% (from 40% to 90% for the impersonal model and 94% to 98% for the personal model) looking at Table I and Table II with 43 features. Other classifiers are also statistically better using normalized features in a number of cases. For 70 number of feature-bank we again used pairwise T-tests between groups of normalized/randomly permuted and unnormalized/non-permuted data at 97%-98% confidence level. Statistically significant results are in boldface shows in all cases again decision trees and SVM classifiers are statistically better when normalized features are used. This increase was as much as almost 10% (from 87% to 97% for the personal model) looking at Table III-IV. Table V-VIII shows single subject cross-validation process for each activity recognition. Examining the overall and single subject cross-validation accuracy of all classifiers, each personal and the impersonal model becomes statistically significant in a number of cases. This leads us to prefer the use of normalization and random permutation in future applications of this method and in practical implementations. All of our chosen classifiers provide acceptable performance but decision tree and SVM are found to be better than other classifiers and have a number of other attractive features to their applicability. In terms of the applicability of the smartwatch system, decision tree and SVM have smaller confidence intervals implying that they have more reliability in their training than other classification models. In terms of the applicability of the smartwatch system, the personal model has smaller confidence intervals implying that personal model has more reliability overall accuracies for both numbers of feature banks.
Looking at Table I-IV, KNN, Decision trees and SVM perform much better in the personal model for both number of features with normalization and for overall accuracy prospect; decision trees outperformed then KNN and SVM. The highest obtained accuracies are in boldface. For further analysis the best-recommended system settings are 70 features normalized and randomly permuted used by Decision tree and SVM classifier to recognize the individual's activity on a smartwatch. Given a fixed amount of data, the performance of KNN and SVM decreases significantly when the window size is increased as it reduces the number of training samples but this does not affect decision tree classifier. Since decision tree, KNN and SVM classifiers trained using 70 normalized and randomly permuted features happens to be the best setting for smartwatch based PAR system. Results are summarized in Fig. 2 which presents the overall classification accuracy with 98% confidence interval. According to these results, decision tree classifier shows higher accuracy for every size of the window. The performance of SVM and KNN through good but is relatively less than that of the decision tree in all cases. Overall, decision trees performance happens to be best followed by SVM as their performance did not much degraded when changing the size of samples per window. On the other hand, the performance of KNN reduced significantly as the window size is increased. We think that it is because we have a fixed amount of data. So a bigger window size results in a smaller number of training samples. To confirm this, we repeated the same experiment for each activity individually. The entire results are summarized in Table I-IV which confirms our hypothesis. The additional experiments on single subject cross-validations are presented in Table V-VIII for both feature banks and personal and impersonal models respectively.
# Conclusion {#sec8}
Feature banks created by different smartphone-based physical activity recognition studies can be used for physical activity recognition on smartwatches because smartphone also uses accelerometer, gyroscope and magnetometer sensors and hence the same set of features is acceptable for a smartwatch. Decision trees, SVM and KNN showed the best results with a minor difference. Additionally, Naive Bayes and ensemble learning with Bag classifiers also produce a good performance. Furthermore, feature normalization and random permutation processes significantly improve the classification performance in several cases. In addition to above, without any doubt, changing window size affects the recognition accuracy. From results, we observed that every activity and each classifier's results of recognition is highly correlated with window size. In general, each classifier performs well when features are extracted using 25-125 samples per window. Personal models perform better for smartwatch based physical activity recognition providing an average accuracy of 98%. Furthermore, the impersonal model produces 94% and 90% accuracy respectively with the same settings of the personal model. Finally, the recommended combination for smartwatch based physical activity recognition system is decision tree classifier trained using 25-125 samples per window of normalized and randomly permuted features. In our future work, we will investigate the effects of several linear/non-linear supervised and unsupervised feature selection methods.
# Introduction {#sec1}
Last few years have witnessed a massive growth in consumer interest in smart devices \[1\]. These include smartphones, smart TVs, and recently smartwatches. Because of their increased penetration, smartwatches have claimed a handsome market share. A typical smartwatch contains a heart rate monitor, GPS, thermometer, camera and accelerometer. Thus, it can provide a variety of services. These services include temperature and pulse measurements and the number of calories consumed when performing a physical activity, such as walking, running, cycling and so forth \[2\]. Recently, smartwatches have emerged in our daily lives and like their smartphone counterparts, smartwatches do contain a gyroscope, magnetometer and accelerometer sensors to support similar capabilities and applications. These applications include health applications i.e. calories consumption that require sensor based physical activity recognition. A number of calories consumed during physical activities vary from one activity to another. Therefore, the first thing to do when estimating the number of calories consumed, is to first recognize which activity have been performed \[3\]. This work explores the idea of physical activity recognition on smartwatches using the embedded gyroscope, magnetometer and accelerometer sensors. The gyroscope, magnetometer and accelerometer sensors are ideal for physical activity recognition. We further discussed the detailed analysis of various feature banks and classification methods to seek the best settings for optimum performance on two different types of models such as impersonal and personal model over the smartwatch. Personal models referred as the classifier is built using the data only form one specific user whereas the impersonal model referred as the classifier is built using the data form every user except the one under study. To summarize, the main contributions of this work is towards answering the following questions: 1): Existing smartphone-based physical activity recognition studies have a number of features that were utilized for smartphone-based physical activity recognition; can they be used for smartwatch-based physical activity recognition as well. 2): Which of the most commonly used classifiers from the reference field provides the best recognition accuracy. 3): Does features permutation help to improve the accuracy. 4): Does features normalization help to improve the performance of the classifier. 5): Which model (Personal or Impersonal) works better for smartwatch-based physical activity recognition. 6): Does changing the window size affect the recognition accuracy. Physical activity recognition in general using motion sensors especially using smartphones has been studied in recent years \[4-6\] and currently being studied extensively \[7-8\] but there are a few studies on physical activity recognition using smartwatches \[9-18\] in which the authors studied the role of the smartwatch in physical activity recognition system for health recommendation purposes. For instance, Guiry et al. \[9\] proposed physical activity recognition system to recognize nine different activities using five different classifiers. These activities include walking, standing, cycling, running, stair ascent, stair descent, elevator ascent and descent. However, the authors studied smartphone and smartwatch based activity recognition separately and did not combine the data from both devices. Furthermore, for real-time data collection, the author's used magnetometer, accelerometer and gyroscope sensors for smartphone and accelerometer for a smartwatch. Trost et al. \[10\] used sensors on hip and wrist to detect physical activities using logistic regression based classification method. In \[10\], Trost et al. just focused on showing the potential of using the wrist position for physical activity recognition for health recommendation system. Chernbumroong et al. \[11\] used a single wrist-worn accelerometer to detect five different physical activities for recommendation system. These activities include standing, sitting, lying, running and walking. However, Da-Silva et al. \[12\] used the same sensor to detect the eight different physical activities for recommendation system. In addition to the activities identified in \[11\] Da-Silva et al. include the activity of working on a computer too. The works \[13\] and \[14\] detect the eating activity using a Hidden Markov Model (HMM) with a wrist-worn accelerometer and gyroscope sensors. Ramos-Garcia identifies the eating activity by dividing it into further sub-activities in \[13\] which includes resting, drinking, eating, using utensil and others. Meanwhile, Dong et al. \[14\] proposed to differentiate eating periods from non-eating periods to identify the eating and not-eating activity. Sen et al. \[15\] used both accelerometer and gyroscope sensor data to recognize the eating and non-eating activities. Furthermore, the authors used smartwatch camera to take images of the food being consumed to analyze what a user is eating. School et al. \[16\] presented a feasibility study to detect and identify the smoking activity using accelerometer sensor where School et al. only reported a user-specific accuracy for smoking activity. Whereas, A. Parade et al. \[17\] used both accelerometer, magnetometer and gyroscope sensors to recognize smoking puffs where they just try to distinguished smoking activity from other activities. A similar work has been reported in \[18\] where the the authors used support vector machine (SVM) to recognize six different physical activities. These activities include standing, walking, writing, jacks, smoking and jogging. Kim et al. \[19\] proposed the collaborative classification approach for recognizing daily activities using smartwatches. The authors exploit a single off-the-shelf smartwatch to recognize five daily activities; namely eating, watching TV, sleeping, vacuuming and showering. For fair comparisons Kim et al. suggested using single and multiple sensors-based approaches to compare with their proposed approach with the overall accuracy of 91.5% with the improved recall rate up to 21.5% for each activity. Weiss et al. showed that accelerometers and gyroscopes sensors that sense user's movements and can help identify the user activity \[20\]. Weiss et al. propose to compare the smartphone and smartwatch-based user activity recognition in which smartwatch have the ability to recognize hand-based activities namely eating which is not possible through a smartphone. They showed that the smartwatch is able to recognize the drinking activity with 93.3% accuracy while smartphones achieve only 77.3%. Nurwanto et al. \[21\] proposed a daily life light sport exercise activities. These activities include squat jump, push up, sit up and walking. Nurwanto et al. focuses only on accelerometer-based signal captured by left-handed persons and the obtained signals were processed through sliding window approach with the k-nearest neighbor method and dynamic time warping as the main classifier. Furthermore, Nurwanto et al. proved that k-NN together with dynamic time warping method is an efficient method for daily exercise recognition with the accuracy of 76.67% for push up 96.69% for squat jump and 80% for sit up activity. They intentionally exclude the walking activity for their experimental process because of the random patterns. Al-Naffakh et al. \[22\] proposed the usefulness of using smartwatch motion sensor namely gyroscope and accelerometer to perform user activity recognition. Al-Naffakh et al. proved the natures of the signals captured are sufficiently discriminative to be useful in performing activity recognition. Different activities can be recognized using motion sensors at the wrist position i.e. drinking, eating, smoking, walking, walking down stairs, walking up stairs, running, jogging, elevator up, elevator down, doing regular exercises, writing etc. However there is no study that discusses these activities all together with the recommendations of optimum system settings. Therefore, to extend the existing works, we evaluate the combination of different feature banks and classification techniques for the reliable recognition of various activities. We also evaluate three different built-in motion sensors. Moreover, unlike the current trends, we also evaluate the effect of window size on each activity performed by the individual user and its impact on the recognition performance of simple and complex activities. Although the effect of window size on physical activity recognition has been studied by Huynh et al. in \[23\] but the windowing effect on complex activities is yet not fully explored. Moreover, unlike the recent works, we explore the use of both *Personal* and *Impersonal* models individually for each activity as explained; 1): Personal model means the classifier is trained using the data only from one specific individual whereas, the Impersonal model means the classifier is built using the data from every user except the one under study. 2): For both models, one subject cross-validation process is adopted. 3): Finally, our study suggests the optimum system settings for any activity recognition system based on smartwatches. This work analyzes the effects of combining gyroscope, magnetometer and accelerometer sensors data for physical activity recognition for different size of windows varying from one second to 12 seconds long window. We further analyze the effect of various sampling rates on the recognition performance in different scenarios. However, in our current study, we still have room for further improvement. For example, our current dataset is not balanced (as shown in Fig. 1) for all performed activities which might lead to the biased results towards majority classes. To handle the said problem to some extent, we used one subject cross-validation process for different models (personal and impersonal) which somehow overcome the class imbalance issue. This study also overcomes the limitations of overlapping while segmenting the raw data which forcefully improve the recognition performance because this overlap cause using some part of data in both training and testing which limiting the use of such techniques in a real-time environment. Moreover, we proposed and suggest some methods which helps to improve the recognition performance for those activities that have low recognition performance. The rest of the paper has the following organization. Section 2 describes the material and methods. Section 3 contains experimental settings. Section 4 presents the experimental results and discussion and finally section 5 concludes the paper with possible future directions.
# Materials and Methods {#sec2}
Today's smartwatches are equipped with a variety of motion sensors that are useful for monitoring device movements like tilt, rotate and shake. Some of these sensors are the ambient light sensor, accelerometer, compass, gyroscope and GPS sensors. For experimental study we only collected the raw signals using three different sensors such as accelerometer, magnetometer and gyroscope sensor.
## Data Collection {#sec3}
We collected the raw signals for five different activities from six users (three female and three male) having a mean age of twenty-five years old. The criterion for selecting the subjects is based on the gender because different genders exhibit different patterns when performing the same activity. These activities include walking, walking upstairs, walking downstairs, running and jogging. All subjects performed these activities twice on each day for more than a month. Therefore we collected raw data from the same user for same activity but performed on different days. The participants enrolled in our study are approved by the laboratory head because this is a formal prerequisite because our experiments involved human subjects and there is a negligible risk of injury. The involved subjects then asked to answer few nontechnical questions like gender, age, height, weight, left or right handed etc, which we used to characterize our study. Then the subjects were asked to fastens the smartwatch on their wrist and places a Bluetooth paired smartphone in their pocket. Both devices run a simple custom designed application that controls the data collection process and instructs the participant to first add their name and then select the activity from the list of five different activities and the sensor from three different sensors. Once the initial instructions were done, turn the smartphone screen off and place the smartphone into the pants pocket. The smartphone instructs the smartwatch running our paired data collection application to collect the raw signal at the 20Hz rate. Each of these sensors generates 3-dimensional signals and appends a time-stamp to the values. Every after five minutes the smartwatch sends the data to the smartphone and after a successful transmission, the smartphone vibrate to notify the user that the data collection process is successfully completed and they can stop the current activity.
## Feature Extraction {#sec4}
The embedded sensor generates three-dimensional time series signals which are highly fluctuating and oscillatory in nature. The oscillation and fluctuations make the physical activity recognition more difficult than other applications in nature. These raw signals are shown in Fig. 1 for individual activity performed by the single user. Therefore, it is compulsory to gather the nontrivial signals from raw data through a feature extraction process. The raw signals are divided into several equal sized windows to control the flow rate and pass fewer data to the system to extract the meaningful information. In \[24\] Khan et al. have presented a detailed analysis of using different kinds of features with a different number of samples per windows for smartphone-based physical activity recognition. From these raw signals, the extracted features are divided into two different feature banks. First feature bank include the features obtained by taking; average, median, variance, standard deviation, interquartile, autocorrelation, partial autocorrelation, coefficients of autoregressive model, coefficients of moving average model, coefficients of autoregressive moving average model and wavelet coefficients. Second feature bank includes the features obtained by taking; average acceleration, average absolute difference, standard deviation, average resultant acceleration and the average difference between peaks. We also calculate the binned distribution in which we determine what fraction of reading fall in a 10 equal-sized bins. This function generates 10 features. Every feature except average resultant acceleration is extracted for each axis. These features are extracted from each axis of the three-dimensional acceleration signal and in total for each window of acceleration data 43 (first-feature-bank) and 70 (second-feature-bank) features are extracted respectively. Prior to the feature extraction process moving average filter of order three is used for noise reduction.
## Classifiers {#sec5}
Khan et al. and Saputri et al. showed that support vector machines (SVM) and artificial neural networks (ANN) are superior to other traditional classification methods for user-independent physical activity recognition system in \[24-26\]. K-nearest neighbor (KNN) is also one of the most popular classification method used for smartphone-based physical activity recognition and user identification \[27\]. ELM \[28\], Naive Bayes \[29\], Adaptive learning/Ensemble learning (BAG) \[30\] and Decision tree \[31\] classifiers are also famous for smartphone-based physical activity recognition system. All these classification methods have shown several advantages in several fields most probably in their respective domains. Once again, given that ours is an exploratory study and the fact that each of these classifiers has shown a good performance in their respective studies. Instead of choosing one we decided to explore the use of all of these classifiers except neural networks in this work.
# Experimental Settings {#sec6}
The following experimental studies are performed using all the data to create both personal and impersonal models. In our first study, using a fixed size of 75 samples per window the performance of five classifiers are compared to the following three cases on both feature banks and for both personal and impersonal models: 1): Randomly permuted features with normalization. 2): Without random permutation and normalization. 3): Randomly permuted features without normalization. In the second study, the best setting for each classifier is chosen and tested multiple times while changing the number of samples in a window *i.e. {25, 50, 75, 100, 125, 150, 175, 200, 225, 250, 275, 300}* samples. The goal is to measure the effect of changing the window size on the performance of each classifier for both personal and impersonal model respectively. In our both studies, the training and testing data is randomly divided and classification results are obtained using 10-fold cross-validation process. The values used for different parameters of classifiers are as follows where all these parameters are carefully tuned and optimized. SVM is trained with quadratic kernel function is used, KNN with Euclidean distance is used and K is set to 10, Ensemble model (BAG) with tree model and the number of weak classifiers are set to 50 and the numbers of decision tree are set to 85.
# Experimental Results and Discussion {#sec7}
This section contains the obtained results for smartwatch-based physical activity recognition system. To validate our system we conduct different experiments. Our first experiment is for both personal and impersonal model-based physical activity recognition analyses. In this experiment, we conduct a detailed comparison of two different feature banks and five classifiers for both personal and impersonal models. Our second experiment explains the physical activity recognition process behavior with a different number of samples per window within the best settings obtained in our first experiment. Finally, the third experiment presents the single subject cross-validation recognition for each activity for both personal and impersonal models individually. All these experiments are based on 10-fold cross-validation process and all these experiments are carried out using Matlab R2014b installed on core i5 and 8GB of RAM machine. In the introduction section; we stated some important research questions which our explanatory study seeks to provide the basis of answers. The answers to those questions based on the obtained results are as follows: Based on our findings, yes the features used in existing smartphone-based physical activity recognition studies can also be used for the smartwatch-based physical activity recognition. If we consider normalized and random permutation feature space, classifiers perform much better opposed to the case of without normalization and random permutation in several cases. Table I-IV present the 98% confidence intervals about overall physical activity recognition rate. Using pairwise T-tests between groups of normalized/randomly permuted and unnormalized/non-permuted data at the 98% confidence level.
Statistically significant results are in boldface, shows decision trees and SVM classifiers are statistically better when normalized and randomly permuted features are used. This increase was as much as almost 50% (from 40% to 90% for the impersonal model and 94% to 98% for the personal model) looking at Table I and Table II with 43 features. Other classifiers are also statistically better using normalized features in a number of cases. For 70 number of feature-bank we again used pairwise T-tests between groups of normalized/randomly permuted and unnormalized/non-permuted data at 97%-98% confidence level. Statistically significant results are in boldface shows in all cases again decision trees and SVM classifiers are statistically better when normalized features are used. This increase was as much as almost 10% (from 87% to 97% for the personal model) looking at Table III-IV. Table V-VIII shows single subject cross-validation process for each activity recognition. Examining the overall and single subject cross-validation accuracy of all classifiers, each personal and the impersonal model becomes statistically significant in a number of cases. This leads us to prefer the use of normalization and random permutation in future applications of this method and in practical implementations. All of our chosen classifiers provide acceptable performance but decision tree and SVM are found to be better than other classifiers and have a number of other attractive features to their applicability. In terms of the applicability of the smartwatch system, decision tree and SVM have smaller confidence intervals implying that they have more reliability in their training than other classification models. In terms of the applicability of the smartwatch system, the personal model has smaller confidence intervals implying that personal model has more reliability overall accuracies for both numbers of feature banks.
Looking at Table I-IV, KNN, Decision trees and SVM perform much better in the personal model for both number of features with normalization and for overall accuracy prospect; decision trees outperformed then KNN and SVM. The highest obtained accuracies are in boldface. For further analysis the best-recommended system settings are 70 features normalized and randomly permuted used by Decision tree and SVM classifier to recognize the individual's activity on a smartwatch. Given a fixed amount of data, the performance of KNN and SVM decreases significantly when the window size is increased as it reduces the number of training samples but this does not affect decision tree classifier. Since decision tree, KNN and SVM classifiers trained using 70 normalized and randomly permuted features happens to be the best setting for smartwatch based PAR system. Results are summarized in Fig. 2 which presents the overall classification accuracy with 98% confidence interval. According to these results, decision tree classifier shows higher accuracy for every size of the window. The performance of SVM and KNN through good but is relatively less than that of the decision tree in all cases. Overall, decision trees performance happens to be best followed by SVM as their performance did not much degraded when changing the size of samples per window. On the other hand, the performance of KNN reduced significantly as the window size is increased. We think that it is because we have a fixed amount of data. So a bigger window size results in a smaller number of training samples. To confirm this, we repeated the same experiment for each activity individually. The entire results are summarized in Table I-IV which confirms our hypothesis. The additional experiments on single subject cross-validations are presented in Table V-VIII for both feature banks and personal and impersonal models respectively.
# Conclusion {#sec8}
Feature banks created by different smartphone-based physical activity recognition studies can be used for physical activity recognition on smartwatches because smartphone also uses accelerometer, gyroscope and magnetometer sensors and hence the same set of features is acceptable for a smartwatch. Decision trees, SVM and KNN showed the best results with a minor difference. Additionally, Naive Bayes and ensemble learning with Bag classifiers also produce a good performance. Furthermore, feature normalization and random permutation processes significantly improve the classification performance in several cases. In addition to above, without any doubt, changing window size affects the recognition accuracy. From results, we observed that every activity and each classifier's results of recognition is highly correlated with window size. In general, each classifier performs well when features are extracted using 25-125 samples per window. Personal models perform better for smartwatch based physical activity recognition providing an average accuracy of 98%. Furthermore, the impersonal model produces 94% and 90% accuracy respectively with the same settings of the personal model. Finally, the recommended combination for smartwatch based physical activity recognition system is decision tree classifier trained using 25-125 samples per window of normalized and randomly permuted features. In our future work, we will investigate the effects of several linear/non-linear supervised and unsupervised feature selection methods. | {'timestamp': '2019-04-23T02:14:29', 'yymm': '1706', 'arxiv_id': '1706.01720', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01720'} |
null | null |
# Introduction and statement of the results {#sec:intro}
We consider dynamical systems defined by the iteration of a meromorphic function \[f: \mathbb C \to \widehat{\C},\] on the complex plane, especially those with an essentially singularity at infinity (transcendental). Motivating examples of such maps are given by root-finding algorithms like, for instance, Newton's method applied to any entire transcendental map. In this setting the Riemann sphere \(\widehat{\C}\) splits into two invariant sets: the *Fatou set* \(F(f)\), which consists of points for which the family of iterates \((f^n)\) is well defined for all \(n>0\) and normal in some neighborhood; and its complement \(J(f)\) known as the *Julia set*. The connected components of the Fatou set or *Fatou components* which are periodic can be completely classified based on the possible limit functions of the family of iterates. Indeed, if \(U\) is a periodic Fatou component, then either \(U\) is a component of a basin of attraction of an attracting or parabolic cycle; or \(U\) is a topological disk (*Siegel disk*) or annulus (*Herman ring*) on which the map is conjugate to an irrational rigid rotation; or the iterates of \(f\) (or multiples of them) tend uniformly on compact subsets to the essential singularity at infinity, in which case \(U\) is a component of a cycle of *Baker domains*. If a Fatou component is neither periodic nor preperiodic, then it is called a *wandering domain*. Neither Baker domains nor wandering components are present in the dynamical plane when \(f\) is a rational map. For classical background on the dynamics of meromorphic functions we refer to the survey or the articles. The set \(S(f)\) of finite *singular values* of \(f\) plays an important role in determining the dynamics of the map. Recall that singular values are either *critical* or *asymptotic values*. Among well known classes of transcendental maps are those with a finite set of singular values (the Speisser class \(\mathcal{S}\)) or with a bounded set of singular values (the Eremenko--Lyubich class \(\mathcal{B}\)). The orbits of the singular values and their accumulation points form the *postsingular set* \[P(f) = \mathbb C \cap \bigcup_{s\in \overline{S(f)}} \bigcup_{n=0}^\infty f^n(s),\] where we neglect the terms which are not defined or are infinite. The importance of singular values lies on the fact that periodic Fatou components are in some way associated to the postsingular set. Indeed, any basin of attraction of an attracting or parabolic cycle must contain a singular value, while the boundary components of Siegel disks and Herman rings are contained in the closure \(\overline{P(f)}\) of the postsingular set. The relation between singular values and Fatou components that are specific for transcendental maps, namely Baker and wandering domains, is less clear. This problem is nowadays one of the challenges in transcendental dynamics, and the object of this paper. It is worth noticing that infinitely many singular values are necessary for these types of components to exist at all, and that entire maps in class \(\mathcal{B}\) do not have Baker domains, nor wandering domains tending to infinity under iteration. Hence, results in this direction must go beyond these classes of functions. It is known that for a wandering domain \(U\), all finite limit functions of \(\{f^n|_U\}\) lie in the derived set of \(P(f)\). Moreover, if for a Baker or wandering domain \(U\) we have \(f^{np}|_U \to \infty\) as \(n \to \infty\) for some \(p \ge 1\), then \(\infty\) is in the derived set of \(\bigcup_{k=0}^{p-1} f^k(S(f))\). These results can be used to rule out the existence of wandering domains in certain situations. The existence of such domains was also excluded for some entire maps studied in and some meromorphic maps with finitely many poles, which are Newton methods of entire functions. There are known examples of Baker domains on which \(f\) is univalent (and hence contains no critical values), has finite degree larger than \(1\) or infinite degree. We refer the reader to for classification and examples of such components. Bergweiler in gave an example of a Baker domain of an entire map lying at a positive (Euclidean) distance from the postcritical set and therefore showing that a very mild relation between these objects is possible. Examples of wandering domains, with the first one given by Baker in, are not so numerous. Most of them are constructed either using approximating theory, by the lifting method (that is by lifting a function with no zeros by the logarithm and therefore converting periodic components into wandering domains), or by quasiconformal surgery. In the first case the method does not allow much control on the postsingular set of the map, while in the second case the relation of the postsingular set to the wandering domain is completely determined by the type of periodic Fatou component of the original map. As an example, take the function considered in, defined as \[f(z)=2-\lambda-\log(2-\lambda) + 2 z-e^z\] with \(\lambda=e^{2\pi i (1-\sqrt{5})/2}\), which is the lift of \(g(w)= \frac{e^{2-\lambda}}{2-\lambda} we^{-w}\) (see Figure [\[siegel\]](#siegel){reference-type="ref" reference="siegel"}). The map \(f\) has infinitely many orbits of simply connected wandering domains on which \(f\) is univalent, while the postsingular set is dense in the boundary of each of these wandering domains. To our knowledge, this is the only explicit example of a simply connected wandering domain on which \(f\) is univalent. All other (lifting) examples contain critical points inside wandering components lifted from basins of attraction of attracting or parabolic cycles. We mention also the inspiring examples by Kisaka and Shishikura in, where they construct wandering domains of eventual connectivity two for maps whose all singular values lie in preperiodic components (see also ). The above examples show different possibilities for the relationship between Baker or wandering domains and the set \(P(f)\). Our goal in this paper is to present some new results in this direction. Our first result concerns Baker domains. In, Bergweiler showed the following result for invariant Baker domains of entire functions.
Here and in the sequel \(\dist\) denotes the Euclidean distance. Additionally, as we mentioned above, Bergweiler provided an example (which inspired the example in Figure [\[siegel\]](#siegel){reference-type="ref" reference="siegel"}), where the postcritical set is located at a positive distance from the Baker domain, showing in some sense the sharpness of the theorem. We remark that Theorem [\[thm:berg\]](#thm:berg){reference-type="ref" reference="thm:berg"} holds for entire functions and therefore all considered Baker domains are simply connected. A version of Theorem [\[thm:berg\]](#thm:berg){reference-type="ref" reference="thm:berg"} for meromorphic transcendental maps (and hence for Baker domains which are not necessarily simply connected) was proven by Mihaljević-Brandt and Rempe-Gillen in, with conclusion (2) replaced by \[\sup_{n\geq 1} \left|\frac{p_{n+1}}{p_n}\right| < \infty.\] They also asked a question, whether the stronger conclusion (3) holds in this setting. In this paper we answer this question positively under a mild additional hypothesis, proving the following.
We remark that the assumption on an unbounded component of \(\mathbb C\setminus U\) (which is equivalent to the condition that \(\{\infty\}\) is not a singleton component of \(\widehat{\C} \setminus U\)) is trivially satisfied if the Baker domain is simply connected. This is always the case if \(f\) is Newton's method of an entire map, as it was shown by the authors in. Theorem A is a consequence of a technical lemma (Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"}) which generalizes and, additionally, has several applications concerning the relation between points in \(P(f)\) and wandering domains. The following consequence of Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"}, although stated for a general Fatou component, is most meaningful when \(U\) is a wandering or Baker domain. In particular, it contributes to an answer to Question 11 from.
The second application of Theorem B concerns functions for which the Julia set and the postsingular set are apart from each other. More precisely, following Mayer and Urbański we consider the class of *topologically hyperbolic* meromorphic functions. This class was also considered by Stallard in in the context of entire maps, where she proved that in this case \(|(f^n)'(z)| \to \infty\) for all \(z \in J(f)\).
This condition can be regarded as a kind of weak hyperbolicity in the context of transcendental meromorphic functions, while hyperbolicity is (usually) defined by the condition that \(\overline{P(f)}\) is bounded and disjoint from the Julia set. In particular, hyperbolic maps are always in class \(\mathcal B\). On the contrary, topologically hyperbolic maps could have sequences of (post)singular points in the Fatou set accumulating at infinity. See e.g. the discussion on various kinds of hyperbolicity for transcendental maps in. Note also that topologically hyperbolic maps do not possess parabolic cycles, rotation domains or wandering domains which do not tend to infinity (otherwise they would have a constant finite limit function in \(\overline{P(f)}\cap J(f)\), see ). Examples of topologically hyperbolic maps include many Newton's methods of entire functions (see Propositions [\[prop:sine_tangent\]](#prop:sine_tangent){reference-type="ref" reference="prop:sine_tangent"} and [\[prop:the_other\]](#prop:the_other){reference-type="ref" reference="prop:the_other"}). The following corollary shows that Baker or wandering domains of topologically hyperbolic maps either contain postsingular points or arbitrarily large disks.
Corollary C may be used in many instances to exclude the presence of wandering domains for a given map. Examples can be found in Section [4](#sec:ex){reference-type="ref" reference="sec:ex"}, where we show that certain Newton maps outside class \(\mathcal{B}\), with infinitely many poles, do not possess wandering Fatou components. In particular, we prove (see Theorem D and Corollary E) that Newton's method of the function \(h(z) = ae^z + b z + c\) has no wandering domains for many parameters \(a, b, c\), including the case when \(a,b,c \in \mathbb R\), \(a,b \neq 0\) and the map \(h\) has a real zero.
# Preliminaries and tools {#sec:prelim}
We use the notation \(\mathbb D(z, r) = \{w \in \mathbb C: |w-z| < r\}\) and \[\mathbb D(A,r) = \bigcup_{z \in A} \mathbb D(z, r)\] for \(z \in \mathbb C\), \(A \subset \mathbb C\), \(r >0\). By \([z_1, z_2]\) we denote the straight line segment connecting \(z_1\) to \(z_2\). The length of a rectifiable curve \(\gamma\) is denoted by \(\ell(\gamma)\). The symbol \(\diam\) denotes the Euclidean diameter. We set \[\dist(z, B) = \inf\{|z-w|: w \in B\}, \quad \dist(A, B) = \inf\{|z-w|: z\in A, \: w \in B\}\] for \(z \in \mathbb C\), \(A, B \subset \mathbb C\). The closure and boundary of a set \(A \subset \mathbb C\) are denoted, respectively, by \(\overline{A}\) and \(\partial A\). We use the following classical results on the distortion of conformal maps (see e.g. ).
In this section we prove the following lemma, which is the main tool in the proofs of our results. It is a generalization of. The estimates are similar to the ones in, which were formulated in terms of the hyperbolic metric.
# Proofs of main results {#sec:A}
# Examples {#sec:ex}
As an application of the above results we will consider some examples of topologically hyperbolic maps for which we can rule out the existence of wandering domains. All the examples are Newton's methods for some transcendental entire functions. Some of the maps considered in this section were studied in detail in . Here we show additionally that none of them has wandering domains. The first proposition was proved previously by Bergweiler and Terglane in by the use of Sullivan's quasiconformal deformations technique. Here we present an elementary proof based on Corollary C.
The second proposition is proved in a similar way.
The main result of this section is to exclude the presence of wandering domains for the following class of maps.
The following corollary shows explicit values of parameters for which the assumptions of Theorem D are satisfied.
To prove Theorem D we will apply Corollary C. Hence, the first step is to show that \(N_h\) is a topologically hyperbolic map. Precisely, we have the following.
This ends the proof of Theorem D. We finish this section proving Corollary E. | {'timestamp': '2017-06-07T02:06:00', 'yymm': '1706', 'arxiv_id': '1706.01732', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01732'} |
null | null |
# [\[sec:level\]]{#sec:level label="sec:level"}Introduction
The two most important methods to obtain reliable quantitative spectral information about the electronic properties of a superconductor are Giaever tunneling and point contact Andreev spectroscopy. In tunnel spectroscopy two metal thin films are weakly coupled by an insulating tunnel barrier, leading to a current-voltage characteristic which is controlled by the unperturbed superconducting densities of states in both metals, \(N_s(E)\), and their occupation, given by the Fermi-functions, \(f_0(E)\). The technique can also be used successfully to study the proximity-effect in superconducting bilayers as experimentally shown by Wolf and Arnold, but requires the difficult development of an opaque tunnel barrier. The second method, point contact Andreev spectroscopy, has become a standard tool to evaluate the microscopic properties of new bulk materials. The experimental configuration consists of a macroscopically sized point-shaped metal wire, which touches a superconducting material, usually a single crystal. In the contact area the conductance in both the superconducting and normal regime is dominated by the channels with the highest transmission usually loosely called 'pinholes'. Thus, there is no need to know the exact nature of the contacting layer and the transmissivity of the point contact can be assumed to reach values in the order of one, without disturbing the properties of the superconductor. This latter assumption is valid because the two bulk materials are connected by an area which is very small compared to the lateral dimensions of the materials and assumed to be smaller than the elastic mean free path both materials (ballistic transport). Such a geometry leaves the reservoirs undisturbed, a crucial condition for the determination of the electronic parameters of the superconductor and generalized in the Landauer-Büttiker picture of quantum transport. Our aim in this paper is to apply Andreev spectroscopy to the proximity-induced superconducting state in a 3D topological insulator (3DTI). The application of Andreev-spectroscopy to low dimensional heterostructures is a much less mature experimental technique than for bulk systems. The point contact has to be lithographically defined and is therefore usually larger than for bulk systems, where accidentally formed pinholes of smaller dimensions dominate the transport. These experimental concerns are exacerbated in the case of spectroscopy on proximity-induced superconductivity, because of the need to use two dissimilar materials and, unavoidably, a complex lithographically structured geometry. In fact, very few successful spectroscopic experiments on proximitized systems have been carried out. One example, on diffusive systems, is by Scheer \(et~al.\) , using mechanical break junctions, an approach that merges bulk point contact behavior with thin films. Recently, Kjaergaard \(et~al.\) have presented results on point contact spectroscopy in the ballistic Al/InAs system, which partially fulfills the experimental requirements. It shows the expected doubling of the quantized conductance steps for point contacts in the highly transmissive regime, but exhibits also, from a spectroscopic perspective, many puzzling results and, additionally, unexpected behavior as a function of the tunable point contact transmissivity. A different geometry was used by Zhang \(et~al.\), also employing a tunable point contact, predominantly in the regime of low transmission. We report on a study of a high quality 3-dimensional topological insulator, epitaxially grown strained HgTe, which is proximitized by a conventional superconductor, niobium. In previous experiments we reported on the observation of a 'missing \(n=1\)' Shapiro step, an indication of an unconventional Josephson effect in 3DTI HgTe based Josephson junctions. The same type of observation was subsequently done in Josephson junctions in a 2D topological insulator showing a sequence of even-only Shapiro steps (up to n=10) and emission at half the Josephson frequency. Both signatures indicate at least a fractional \(4\pi\)-periodic Josephson effect, and point towards the presence of gapless Majorana-Andreev bound states. Since the Josephson effect arises from the proximity-induced superconducting state, we are interested in a determination of the energy dependent properties of this induced superconducting state, which in principle serves as a coherent reservoir for the Josephson effect, analogous to the established proximity-effect based niobium superconductor-insulator-superconductor (SIS) junctions. It is crucial to be able to measure these electronic states directly, in particular because the Josephson-effect itself contains only information about the phase difference and the nature of the current-phase relation, but not about its energy dependence. For this reason we designed an experiment which is based on a NcS\(_\textrm{p}\) point contact to emulate Andreev-spectroscopy of the induced superconducting state (N is a normal reservoir, which in our case is a topological insulator, c is the constriction, and S\(_\textrm{p}\) is the proximity-induced superconductor), as schematically shown in Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}a). Therefore, the strained HgTe is defined lithographically to a finite sized bar and covered over a small distance by a conventional superconductor S\(_\textrm{m}\). We assume that an induced superconducting state exists underneath the superconducting material, which we label S\(_\textrm{p}\). The electronic states in this region are the source for the observed Josephson effect. Note that in such a geometry no Majorana zero modes are expected to emerge due to the lack of confinement but unconventional superconducting correlations might be observable. We find that the electronic transport between the N-reservoir and the S\(_\textrm{m}\) reservoir is governed by two energy scales which we identify as the superconducting gap of the niobium film \(\Delta_{\textrm{Nb}}\) and the induced gap in the surface states of the HgTe, labeled \(\Delta_{\textrm{p}}\). By using modeling as introduced by Blonder \(et~al.\) we are able to show that the transmissivity at the HgTe/Nb interface is rather low. We argue that the voltage-carrying state, needed to obtain spectroscopic information, leads to a non-equilibrium occupation of the proximity-induced superconducting state, rendering the device into different experimental conditions, depending on the bias voltage.
# [\[sec:level2\]]{#sec:level2 label="sec:level2"}Sample description
The NcS junctions in this work are based on epitaxially grown layers of strained HgTe sandwiched between Hg\(_{0.3}\)Cd\(_{0.7}\)Te capping layers. These additional layers have a conventional band structure and protect against surface oxidation, which reduces the carrier mobility. They also protect the strained HgTe during subsequent lithographic processing. The HgTe sandwich is shaped as a \(1\,\)m wide bar which at one or both ends tapers out with an angle of about \(45^\circ\). The top Hg\(_{0.3}\)Cd\(_{0.7}\)Te capping layer is removed by dry etching and subsequently covered by niobium, which is in contact with the strained HgTe. Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}a) shows a schematic drawing of the device. The orange part is the source-superconductor, S\(_\textrm{m}\), made of niobium and the blue part is the strained HgTe. At the interface we allow for a finite transmission coefficient which is labeled \(Z_{\textrm{m}}\). This dimensionless barrier is in general connected to the normal state transmission by \(t=(1+Z^2)^{-1}\). The superconducting correlations are induced in the HgTe indicated by yellow dots. The tapered part of the HgTe, not covered by the niobium, is left capped by the Hg\(_{0.3}\)Cd\(_{0.7}\)Te layer, and we assume that this part has the same mobility as the starting material. At the constriction we allow for an additional elastic scattering parameter \(Z_{\textrm{p}}\). The quality of the HgTe layers is characterized using a Hall bar fabricated from the same wafer. At zero gate voltage (\(V_g=0\)) a density of \(n_{\textrm{2D}} \approx 5 \times 10^{11} \, \textrm{cm}^{-2}\) and mobilities of \(\mu \approx 200 \, 000 \, \textrm{cm}^2 /\textrm{Vs}\) are routinely achieved resulting in a mean free path \(l_{\textrm{mfp}}\approx 2-3 \,\)m. The mobility is about ten times lower when tuning the device into the p-regime. As shown in detailed magnetotransport studies, clear quantum Hall plateaus are observed indicating transport mediated predominately by two dimensional states which were shown to originate from the topological surface states. The point contact is fabricated using electron beam lithography and PMMA resist. As HgTe is sensitive to temperatures above \(90 ^{\circ}\)C, all bake-out and lift-off procedures are carried out well below this temperature. In a first step the HgTe mesa is defined using low energy argon sputtering. During this process a thin titanium etch shield, separated by a SiO\(_2\) sacrificial layer from the HgTe, is protecting the mesa. The shield is afterwards removed by a buffered oxide etch dip. The dimensions of the mesa as shown in Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}b) and c) are chosen such that the orifice (\(a=1 \,\)m for Device 1, 2 and 4 and \(a=0.6 \,\)m for Device 3, respectively) is smaller than the ballistic mean free path of the surface states. The size of the normal reservoir is much larger than this length scale, to allow full energy relaxation in this region. In a next step, the superconductor is deposited. Since the interface is buried the cap layer needs to be removed, which is done by argon etching, followed by in-situ magnetron sputtering of about \(110 \,\)nm of niobium. After this the leads for the Ohmic contacts are defined and \(50\,\)nm AuGe/\(50\,\)nm Au is deposited. The contact resistances are usually small (\(<50 \, \Omega\)). To allow control of the charge carrier density in the 3DTI a top gate electrode is evaporated on top of the HgTe (c.f. dashed lines in Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}b), as follows. First, a thin HfO\(_2\) insulator is grown at about a temperature of \(35^{\circ}\)C via atomic layer deposition, followed by the deposition of \(5\,\)nm Ti /\(150\,\)nm Au. Using the same insulator on reference Hall bar structures it is possible to tune the density from \(1 \times 10^{12} \, \textrm{cm}^{-2}\) n-type regime to \(-1 \times 10^{12} \, \textrm{cm}^{-2}\) p-type dominated conductance. A false color SEM picture of a final device without an applied gate is shown in Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}c). For the transport studies the samples are then cooled down in a dilution refrigerator with a base temperature of \(30 \,\)mK (Device 1) or \(120\,\)mK (Device 2-4) and the differential conductance \(dI/dV\) is measured using low excitation and low-frequency lock-in techniques combined with DC measurements as depicted in Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}b). Several devices made from different wafers with and without a top-gate have been measured, yielding all very similar results from which four exemplary devices are discussed.
# Method of analysis
In the design of the experiment, we anticipate that the transport from N to S will be controlled by the process of Andreev reflection, which allows using the theory of Blonder \(et~al.\) (BTK-theory). This theory assumes thermal equilibrium for the relevant states. In the experimental configuration used by us the occupation of states will potentially deviate from the equilibrium Fermi-Dirac distribution. As shown in Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}a) we define three sections through which the transport occurs in our device. In that drawing, the wide uncovered part of the 3DTI constitutes the N side and fulfills the criterion of a proper Landauer-Büttiker equilibrium reservoir with a Fermi-function at the bath temperature \(T_b\) and a Fermi-level \(\mu_\textrm{N}=\mu_{\textrm{p1}}-eV_{\textrm{SN}}\), which depends on the applied applied bias \(V_{\textrm{SN}}\). On the other side of the constriction, located at \(Z_{\textrm{p}}\), the main superconductor S\(_\textrm{m}\), niobium, induces superconducting correlations in the 3DTI bar S\(_\textrm{p}\). Both superconductors form the same macroscopic quantum state. The current through the sample, assumed to enter from the N-part is carried away as a supercurrent. Therefore, we do not expect a voltage drop beyond \(Z_{\textrm{p}}\) and the superconducting side is initially, for zero applied applied bias \(V_{\textrm{SN}}=0\), characterized by an equilibrium Fermi-function \(\mu_{\textrm{p1}}=\mu_{\textrm{p2}}\) at the bath temperature \(T_b\). With this starting point we anticipate that the conductance as a function of voltage \(V_{\textrm{SN}}\) will, in principle, described by: \[\label{BTK} \begin{aligned} I_{SN}&=&\frac{1}{eR_N} \int_{-\infty}^{+\infty} (f_0(E-eV_{\textrm{SN}},T)-f_0(E,T))\\ &&[1+A(E,Z)-B(E,Z)] \mathrm{dE}, \end{aligned}\] where \(f_0(E,T)\) is the Fermi-Dirac distribution at energy \(E\) and temperature \(T\). \(A(E)\), and \(B(E)\) are the probability amplitudes for Andreev and normal reflection of an incident electron from and to the normal reservoir. The normal state resistance \(R_N\) is assumed to be the resistance arising from the finite number of modes carried by the cross-section. The voltage drop is located at the orifice with elastic scattering parametrized by \(Z=Z_{\textrm{p}}\) as indicated in Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}a). We do not know the coefficients \(A(E)\) and \(B(E)\) *a priori*. They contain the spectral information we are interested in and are the result of the interaction of the superconductor with the confined bar of the 3DTI with its limited geometry, finite elastic mean free path and finite interfacial transparency \(Z_{\textrm{m}}\) . In addition it needs to be considered, that the normal part is a 3DTI, where helical surface states dominate the transport. In the covered TI bar we allow for a finite paring-potential \(\Delta\), which implies, that the self-consistency equation of the Bogoliubov-De Gennes equations \[\label{SelfConsistency} \Delta(\vec r)=V_N\sum_{E>0} v^*(\vec r)u(\vec r)[1-2f_0(E)].\] needs to be fulfilled. The value \(\Delta\) depends on the distribution function, which for a driven system may differ from the one assumed for equilibrium reservoirs. Hence, we will analyze our data under the assumption that Andreev reflection, due to a finite value of \(\Delta\), takes place at \(Z_{\textrm{p}}\), which allows us to apply Eq. ([\[BTK\]](#BTK){reference-type="ref" reference="BTK"}) to our system with initially, for low voltages, the equilibrium reservoirs taken to be in the normal side and in the proximitized HgTe on the superconducting side with a finite value of \(\Delta=\Delta_{\textrm{p}}\), although it does not necessarily resemble a BCS like density of states.
# [\[sec:level3\]]{#sec:level3 label="sec:level3"}Experimental data and interpretation
Fig. [\[Fig:fig1\]](#Fig:fig1){reference-type="ref" reference="Fig:fig1"}d) gives an overview of the differential conductance across the point contact for four different devices at zero applied gate voltage and zero magnetic field. At voltages \(\lvert V_{\textrm{SN}}\rvert>1.5\,\)mV, larger than \(\Delta_{\textrm{Nb}}\), the differential conductance is almost constant and a normal state resistance of \(R_{\textrm{N}} = 160-240 \, \Omega\) is observed, depending on the measured device. For voltages around \(V_{\textrm{SN}} \simeq 1.1\,\)meV the conductance is slightly enhanced which is indicated by the black arrows and then starts to decrease for smaller voltages. Close to zero bias, the conductance enhances again resulting in a double peak structure around \(V_{\textrm{SN}}=0\), with a peak separation of about \(100\,\)V for Device 1, and slightly different for the other devices. The red arrows are used to draw the attentions to a sample dependent sub gap feature. The four devices differ with respect to the shape and length of the HgTe bar underneath the superconductor. Device 1 is symmetric with width \(w=1\,\)m and two open ends. Device 2 has a step like shape with partially width \(w\) and partially width of \(0.6\,\)m. Similarly, Device 3 but with the wide 'normal' electrode connected to the wide part rather than the more narrow part. Finally, Device 4 is terminated half-way and implies a largely closed HgTe bar. At present it is not clear whether this should be interpreted as a feature in the relevant non-equilibrium distribution entering Eq. ([\[BTK\]](#BTK){reference-type="ref" reference="BTK"}) or as reflecting a finite size effect of the HgTe in the spirit of the analysis of Kopnin and Melnikov. Systematic shape-dependent experiments are needed to map and evaluate this dependence accurately and to test the full hypothesis. An asymmetric background for negative and positive bias is observed in all devices. The data can be normalized by multiplying with the normal state resistance \(R_N\) measured at \(T>T_c\), as shown in Fig. [\[Fig:tempdependence\]](#Fig:tempdependence){reference-type="ref" reference="Fig:tempdependence"}, to eliminate this slope. We will discuss the observed behavior now in more detail.
## Low voltage data: proximity-induced order parameter
Close to zero bias, we find a strongly enhanced conductance with a double-peak structure in Devices 1-3 and a single peak for the closed bar (Device 4). As shown in Fig. [\[Fig:tempdependence\]](#Fig:tempdependence){reference-type="ref" reference="Fig:tempdependence"}, this double-peak structure merges at higher temperatures to a bell-shaped curve. The data in this figure is taken for Device 1, which we will focus on for a detailed analysis. The conductance is normalized with the normal state resistance \(R_N\) above the critical temperature \(T>T_c\). From Fig. [\[Fig:tempdependence\]](#Fig:tempdependence){reference-type="ref" reference="Fig:tempdependence"}a) it is clear that at \(4.2\,\)K an energy gap opens up, which is on the scale of the superconducting niobium gap. Upon lowering the temperature a peak emerges around \(V_{\textrm{SN}}=0\), which splits in two below \(500\,\)mK. Panel b) of Fig. [\[Fig:tempdependence\]](#Fig:tempdependence){reference-type="ref" reference="Fig:tempdependence"} shows the conductance measured at \(30\,\)mK for increasing values of magnetic field applied perpendicular to the sample. We verified that the response is independent of the direction of the applied magnetic field. For clarity, a small vertical shift has been removed in the presentation of the data in Fig. [\[Fig:tempdependence\]](#Fig:tempdependence){reference-type="ref" reference="Fig:tempdependence"}b) to highlight that the high voltage part of the conductance is immune to these magnetic field strengths. Evidently, the central peak can be suppressed completely by applying a magnetic field. We attribute this central bell-shaped peak, which evolves into a two peak structure as a manifestation of the proximity-induced superconducting order parameter as given by Eq. ([\[SelfConsistency\]](#SelfConsistency){reference-type="ref" reference="SelfConsistency"}). For \(V_{\textrm{SN}}=0\) the system is in equilibrium and the apparent \(\Delta\) is the result of electrons in the HgTe bar underneath the niobium film, which are confined in a certain width and length. Their occupation is given by a Fermi-Dirac distribution and it is part of the superconducting equilibrium reservoir S\(_\textrm{p}\). For finite voltage bias, the current in S\(_\textrm{p}\) is carried away as a supercurrent, and, importantly, the voltage-drop occurring at the interface indicated by \(Z{\textrm{p}}\) in Fig. [\[Fig:two_barriers\]](#Fig:two_barriers){reference-type="ref" reference="Fig:two_barriers"}a), is due to the difference in electrochemical potentials between N on the left of \(Z_{\textrm{p}}\) and S\(_{\textrm{p}}\) on the right of \(Z_{\textrm{p}}\). The scale of the relevant Sharvin resistance is controlled by the number of modes at the \(Z_{\textrm{p}}\) location and by the unknown value of \(Z_{\textrm{p}}\). Therefore, the normalization on \(R_N\) as defined above is not viable in this equilibrium regime. In Fig. [\[Fig:BTKZ1\]](#Fig:BTKZ1){reference-type="ref" reference="Fig:BTKZ1"}a), data for different temperatures are compared with standard BTK-modeling using Eq. ([\[BTK\]](#BTK){reference-type="ref" reference="BTK"}) (cyan) and the model that explicitly takes the surface states of a 3DTI into account from Ref. (magenta) both leading to very reasonable agreements. In this figure we have renormalized the data differently. We have chosen the conductance value at the edge of the gray zone in Fig. [\[Fig:BTKZ1\]](#Fig:BTKZ1){reference-type="ref" reference="Fig:BTKZ1"}b), as a reasonable approximation to the real value of \(R_N\) entering Eq. ([\[BTK\]](#BTK){reference-type="ref" reference="BTK"}). From the comparison shown in Fig. [\[Fig:BTKZ1\]](#Fig:BTKZ1){reference-type="ref" reference="Fig:BTKZ1"}a), we conclude that we find a proximity-induced order parameter \(\Delta_{\textrm{p}}= 70\,\)eV for both models. The fits using Eq. ([\[BTK\]](#BTK){reference-type="ref" reference="BTK"}) were obtained with a small barrier height \(Z_{\textrm{p}}=0.4\). Here, we have assumed that the proximity-induced order parameter \(\Delta_{\textrm{p}}\) leads to a standard BCS like behavior of the coefficients \(A(E)\) and \(B(E)\) as a function of energy and that the normal state is described by a parabolic band dispersion. The model might therefore not capture the microscopic details but makes it suitable to compare to other systems. The treatment of Ref. models the conductance of a NS junction on the surface of a 3DTI, exactly as appropriate for our experiment. The contact between the normal region and the induced superconducting reservoir is modeled as a square potential barrier, where the dimensionless barrier strength \(Z_{\textrm{p}}\) is defined as the product of the barrier height and width. The sub-gap tunnel conductance of the NS junction is then an oscillatory function of the barrier strength \(Z_{\textrm{p}}\) and minimum for values \(Z_{\textrm{p}}=(n+1/2)\pi\), with \(n\) an integer. By applying this model to our experimental data, a rather large barrier can be used. The enhanced conductance can then be seen as a signature of the helical surface states where highly transparent modes are always expected due to Klein tunneling. We interpret the low voltage data as a probe of the induced superconducting state in the 3DTI of strained HgTe. There is no reason to expect *a priori* a s-wave order parameter. In fact we expect deviations, such as for example predicted by Burset \(et~al.\). Since the actual spectra depend on several parameters, a larger data-set is needed to provide a reliable analysis to show the influence of the helical Dirac nature of the surface states. Nevertheless, this open question does not affect the conclusion that we can draw with respect to the identification of the regime, where spectroscopy of the induced superconducting state can reliably be performed.
## High voltage data: niobium order parameter
For voltages larger than \(0.5\,\)meV, the conductance curves in Fig. [\[Fig:tempdependence\]](#Fig:tempdependence){reference-type="ref" reference="Fig:tempdependence"}b) all superimpose, if we excerpt the central part interpreted as the proximity-induced order parameter. The data outside the central part can no longer be interpreted as the conductance of a NcS point contact at \(Z_{\textrm{p}}\). The electronic states in the HgTe bar underneath the niobium are no longer correlated as expressed in Eq. ([\[SelfConsistency\]](#SelfConsistency){reference-type="ref" reference="SelfConsistency"}). For increasing voltage at the location \(Z_{\textrm{p}}\), higher energy quasiparticles are injected into the HgTe bar as depicted in Fig. [\[Fig:two_barriers\]](#Fig:two_barriers){reference-type="ref" reference="Fig:two_barriers"}b). They cannot escape into an equilibrium reservoir because of the large gap of the superconductor niobium and the fact that Andreev reflections do not exchange heat. Therefore, \(f_0(E)\) in Eq. ([\[SelfConsistency\]](#SelfConsistency){reference-type="ref" reference="SelfConsistency"}) becomes a non-equilibrium distribution with relatively hot electrons, which leads in general to a destruction of the proximity-induced order parameter \(\Delta_{\textrm{p}}\), in the same way as a small magnetic field quenches this induced superconducting state. Hence, beyond a voltage of about \(0.5\,\)meV the system has changed and we are left with a non-superconducting HgTe bar in contact with niobium (as shown in Fig. [\[Fig:two_barriers\]](#Fig:two_barriers){reference-type="ref" reference="Fig:two_barriers"}c) with an interface with an unknown transmissivity parametrized by \(Z_{\textrm{m}}\). The change in conductance around \(1.1\,\)mV is now naturally attributed to the superconducting gap of the niobium film. The conductance increases slightly, as expected at the superconducting gap edge. For smaller voltages the conductance reduces, an indication of dominant normal reflections over Andreev reflections (\(B/A>1\)). As shown in Fig. [\[Fig:BTKZ1\]](#Fig:BTKZ1){reference-type="ref" reference="Fig:BTKZ1"}b), we are able to achieve fairly good qualitative agreement with a BTK-analysis as well for this outer gap, using a quite large barrier \(Z_{\textrm{m}} = 1.1\) and \(\Delta_{\textrm{Nb}} = 0.8\,\)meV, indicating a relatively low transparency of the Nb/HgTe interface. We also need to use a relatively large broadening parameter \(\Gamma=0.7\Delta_{\textrm{Nb}}\) which could be caused by the large contact area and spatial gradients at the Nb/HgTe interface.
# [\[sec:level5\]]{#sec:level5 label="sec:level5"}Gate dependence of the conductance
The previous data are all obtained on the electron side (n-type) in which the mobility is high. In Fig. [\[Fig:gate\]](#Fig:gate){reference-type="ref" reference="Fig:gate"}a) conductance data are shown for different gate voltages from \(+1~V\) to \(-4~V\) in which the 3DTI changes from n-to p-type conduction. The curves are normalized to the resistance \(R_N(T>T_c)\) for each gate voltage individually. The behavior of the normal state resistance of Device 2 versus the gate voltage is comparable to the reference Hall-bar where we are able to tune the density from initially n-doped, over the charge neutrality point into the hole dominated regime. We distinguish two regimes from \(1\,\)V to about \(-1\,\)V the device is in the n-conducting regime. In this regime the mobility is high and the point contact is expected to be ballistic. By tuning into the p-regime the mobility reduces by about a factor of ten and the mean free path is now smaller than the size of the point contact and, therefore, expected to be in the diffusive regime. From the conductance curves (Fig. [\[Fig:gate\]](#Fig:gate){reference-type="ref" reference="Fig:gate"}a) it is clear that we observe no longer a signature of the niobium pairing potential in the p-regime. Upon changing the gate voltage, features at the scale of the niobium gap disappear upon approaching the Dirac point (at \(-2.2~V\)). The only significant voltage-dependent feature is around \(\pm~ 100\,\)eV. We assume that this observation is a signature that the NcS point contact is probing the induced superconducting state of the HgTe bar in a diffusive proximity-system, leading to a mini-gap. The height of the zero bias anomaly as a function of gate voltage is quantified using Fig. [\[Fig:gate\]](#Fig:gate){reference-type="ref" reference="Fig:gate"}a), by defining \(dI/dV_{T=30\,\textrm{mK}}-dI/dV_{T>T_c}\), and plotted in Fig. [\[Fig:gate\]](#Fig:gate){reference-type="ref" reference="Fig:gate"}b) as red dots. The amplitude is several tens of \(e^2/h\) in the n-conducting regime and decreases continuously up to the the maximum in resistance region where it saturates at a value of 1-2 \(e^2/h\) depending on the sample.
# General remark about our analysis
The analysis of our data has lead us to discuss the conductance data resulting from the transport through three different electron systems (N, S\(_\textrm{p}\) and S\(_\textrm{m}\)), separated by two interfaces of transparency \(Z_{\textrm{p}}\) and \(Z_{\textrm{m}}\). Following Beenakker it is assumed that any contact between a normal reservoir and a superconducting reservoir is given by \[\label{BeenakkerBTK} G_S=2G_0\frac{G_N^2}{(2G_0-G_N)^2}\] with \(G_0=2e^2/h\) the quantum unit of conductance, \(G_N\) the conductance in the normal state, and \(G_S\) the conductance with one of the electrodes superconducting. This expression is the zero-voltage limit of the classical BTK-formula for different values of transmissivity \(Z\). In order to calculate \(G_S\), often the conductance at \(V>\Delta_s\) is used as \(G_N\) (see also Fig.[\[Fig:gate\]](#Fig:gate){reference-type="ref" reference="Fig:gate"}c)) and implying that this experimental value is independent of the applied bias. The most important implication in our case is that one measures at high voltages not a proximity-induced superconducting gap, but rather the parent superconductor. We suggest that the low voltage data should be understood by acknowledging that the scattering region and the equilibrium reservoirs at \(V_{\textrm{SN}}=0\) should be defined differently from the one at higher voltages, such as in our case \(V>0.8\,\)meV. This distinction is in general not specific to our case but should apply to other topological systems, for example the one studied in Kjaergard \(et~al.\) and Suominen \(et~al.\) and might explain deviations from expected behavior in these two papers.
# Conclusions
In conclusion, we have carried out transport spectroscopy of the proximity-induced pair-potential of a niobium covered bar of strained HgTe, which has been demonstrated to be prone to be a 3DTI. In analyzing the data we allow for a finite pairing potential in the strained HgTe, in contrast to a commonly made quantum transport simplification as introduced by Lambert and Beenakker, in which the properties are assumed to be controlled exclusively by the scattering in the structure. In addition, we take into account how to identify the relevant distribution function over the energies, implying the relevance of a non-equilibrium distribution function in analyzing the data. These results are an important step towards a better understanding and engineering of topological superconductivity and may serve as a building block for a further analysis of the \(4\pi\)-Josephson effect as reported in Refs.. We like to thank A. Akhmerov, W. Belzig, F.S. Bergeret, and B. Trauzettel for many helpful discussions. The work at Würzburg was supported by was financially supported by the German Research Foundation DFG via SFB 1170 "ToCoTronics "and the SPP 1666, the Land of Bavaria (Institute for Topological Insulators and the Elitenetzwerk Bayern and the European Research Council (advanced grant project 3-TOP and 4-TOPS). TMK, who acknowledges support from the European Research Council Advanced grant no. 339306 (METIQUM), by the Ministry of Education and Science of the Russian Federation, contract 14.B25.31.0007 of 26 June 2013 and TMK and EB thank the Alexander von Humboldt Stiftung. P.B. acknowledges support from the European Union's Marie Skłodowska-Curie grant agreement No. 743884. | {'timestamp': '2017-10-20T02:05:13', 'yymm': '1706', 'arxiv_id': '1706.01638', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01638'} |
# Introduction
Quasi-periodic pulsations (QPPs) are usually observed in the light curves of solar flares. They typically display regular and periodic peaks from the total flux based on the time-series analysis. The observations of QPPs can cover almost all the wavelengths, such as soft/hard X-rays (SXR/HXR), extreme-ultraviolet (EUV/UV) and radio emission. In addition, QPPs can be detected with spectral observations, i.e., Doppler shifts, line widths and intensities. So far, it is still unclear which mechanism could cause the QPPs. Previous findings show that QPPs could be interpreted as the MHD wave or produced by the quasi-periodic magnetic reconnection. These QPPs models only apply to one or several events, but fail to explain all the observational features. The observed periods of QPPs range from tens of seconds to tens of minutes. Sometimes, QPPs in the same event could show multiple periods in radio, H\(\alpha\), EUV, SXR, and HXR bands. The period ratio of QPPs in the same event depends on the magnetohydrodynamic (MHD) mode. The ratio equals or almost equals two for the weakly dispersive MHD modes, i.e., torsion, slow magneto-acoustic, and kink. However, it deviates from two for the highly dispersive modes. The difference the period ratio from two is also attributed to additional physical effects like siphon flows, the loop expansions, or the longitudinal density stratifications. QPPs with multiple periods can be detected during all phases of solar flare, i.e., from impulsive to decay phases. However, the fact that the dominant period of flare QPPs becomes longer with the increasing photon energy during the impulsive phase has not been reported yet. Using the high temporal resolution observations from *Fermi* Gamma-ray Burst Monitor (GBM), we explore the QPPs behaviors in a solar flare on 2014 September 10. The new observational finding may be a challenge to the QPPs theory, and could help to refine the model of flare QPPs.
# Observations
In this paper, we investigate an X1.6 flare which takes place in NOAA AR 12158 on 2014 September 10. It is accompanied by a fast coronal mass ejection (CME). It starts at about 17:21 UT, and then reaches its maximum at around 17:45 UT from the *GOES* SXR flux at 1.0\(-\)`<!-- -->`{=html}8.0 Å and 0.5\(-\)`<!-- -->`{=html}4.0 Å, respectively, as shown in Figure [\[flux\]](#flux){reference-type="ref" reference="flux"} (a). Figure [\[flux\]](#flux){reference-type="ref" reference="flux"} (b) shows the light curves at four HXR channels from *Fermi*/GBM, such as 12.0\(-\)`<!-- -->`{=html}27.3 keV, 27.3\(-\)`<!-- -->`{=html}50.9 keV, 50.9\(-\)`<!-- -->`{=html}102.3 keV, and 102.3\(-\)`<!-- -->`{=html}296.4 keV. To display the HXR light curves clearly, we have multiplied a factor for every HXR light curve. They are observed by n2 detector, whose direction angle to the Sun is relative stable during the impulsive phase of solar flare, i.e., between 17:21 UT and 17:45 UT. However, the other detectors frequently change their direction angles. After the impulsive phase (i.e., 17:45 UT), the n2 detector also shifts its direction angle frequently, this results into the HXR peak around 17:47 UT, which may be not real. It is missing data after 17:54 UT. The time cadence of *Fermi* light curves is about 0.256 s, but it automatically becomes to 0.064 s when the flare bursts. Therefore, the light curves from *Fermi* in Figure [\[flux\]](#flux){reference-type="ref" reference="flux"} (b) are interpolated into an uniform cadence, i.e., 0.256 s.
# Data reduction technique
To look closely the periods of QPPs, we decompose every light curve into slowly varying signal and rapidly varying signal. The slowly varying signal is obtained by smoothing (SMOOTH.PRO in IDL) the original data with a timescale boxcar, as shown by the dashed profile overlapping on the solid profile in Figure [\[flux\]](#flux){reference-type="ref" reference="flux"} (b). Previous observation has been suggested that this X1.6 flare may exhibit QPPs with short time periods in HXR bands, i.e., 30 s and 45 s. However, these short timescale vibrations have much smaller amplitude than the strong background emission, which make them difficult to be detected. Therefore, the window timescale of 30 s is chosen to highlight the short timescale fluctuations. The rapidly varying signal is achieved by subtracting the slowly varying signal from its original light curve. And then the rapidly varying signal is normalized by its slowly varying signal, so that it could exhibit the QPPs behaviors much better. Figure [\[qpp1\]](#qpp1){reference-type="ref" reference="qpp1"} (a) shows the normalized rapidly varying signal at HXR 12.0\(-\)`<!-- -->`{=html}27.3 keV between 17:22 UT and 17:38 UT from *Fermi* observations. It clearly shows the periodic peaks and could be identified as QPPs. Note that we only select some time-series flux during the impulsive phase to avoid the frequently changing of the direction angle of n2 detector.
# Results of analysis
Figure [\[qpp1\]](#qpp1){reference-type="ref" reference="qpp1"} (b) shows the wavelet power spectrum from the normalized rapidly varying signal at lower energy channel, i.e., 12.0\(-\)`<!-- -->`{=html}27.3 keV, the red contours outline the significance levels of 99.99%. It displays the dominant period with a broad band between 17:25 UT and 17:29 UT, as for most QPPs in microwave and HXR bands. To further confirm the periodicity, we perform a periodogram analysis for the normalized rapidly varying signal with the Lomb-Scargle periodogram method, as shown in panel (c). The red vertical line represents the 99.99% confidence level defined by. It shows two peaks above the confidence level, and both them have narrow widths. Therefore, the dominant period is obtained from the peak value in the Fourier periodogram, and the error bars are calculated from both the width of the spectral peak in the periodogram and the discreetness of the Fourier frequency grid. Thus two dominant periods within an error bar can be estimated from 17:25 UT to 17:29 UT at 12.0\(-\)`<!-- -->`{=html}27.3 keV, which are 27\(\pm\)`<!-- -->`{=html}2.8 s and 37\(\pm\)`<!-- -->`{=html}3.5 s. The same results can be found at 27.3\(-\)`<!-- -->`{=html}50.9 keV during almost the same time intervals, as seen in Figure [\[qpp2\]](#qpp2){reference-type="ref" reference="qpp2"}. Although the wavelet spectrum still displays weak signal of periodicity after 17:29 UT (panel b), the FFT spectrum only shows two peaks within narrow widths greater than the confidence level at about 27 s and 37 s, which are similar to that in Figure [\[qpp1\]](#qpp1){reference-type="ref" reference="qpp1"} (c). Thus we could derive the periods of 27\(\pm\)`<!-- -->`{=html}2.8 s and 37\(\pm\)`<!-- -->`{=html}3.5 s at 27.3\(-\)`<!-- -->`{=html}50.9 keV between 17:25 UT and 17:29 UT. With the growth of HXR photon energies, the period of QPPs changes obviously to the larger value, as shown in Figures [\[qpp3\]](#qpp3){reference-type="ref" reference="qpp3"}. At medium-energy channel (50.9\(-\)`<!-- -->`{=html}102.3 keV), the QPPs display multiple periods within a broad band from 17:26 UT to 17:33 UT. The shorter periods appear at the beginning of flare impulsive phase, such as between 17:26 UT and 17:29 UT. And the longer period appears at the end of flare impulsive phase, i.e., between 17:28 UT and 17:33 UT (Figure [\[qpp3\]](#qpp3){reference-type="ref" reference="qpp3"} b). Considering the peaks within narrow width in FFT spectrum, we can get three dominant periods within an error bar at 50.9\(-\)`<!-- -->`{=html}102.3 keV, they are 27\(\pm\)`<!-- -->`{=html}1.8 s, 46\(\pm\)`<!-- -->`{=html}4.2 s, and 60\(\pm\)`<!-- -->`{=html}5.8 s, as shown in Figure [\[qpp3\]](#qpp3){reference-type="ref" reference="qpp3"} (c). At higher energy channel (102.3\(-\)`<!-- -->`{=html}296.4 keV), we apply a wider boxcar window of \(\sim\)`<!-- -->`{=html}100 s to smooth the light curve. Figure [\[qpp4\]](#qpp4){reference-type="ref" reference="qpp4"} (a) shows the normalized rapidly varying signal. The wavelet spectrum exhibits a period with a broad band between 17:28 UT and 17:33 UT, as shown in panel (b). However, the FFT spectrum only has one peak with a narrow width that greater than the confidence level, which is around 80 s. Based on these facts, we could obtain the dominant period within an error bar of 80\(\pm\)`<!-- -->`{=html}11 s at higher energy channel. Even if using the short window timescale (\(\sim\)`<!-- -->`{=html}30 s), we could not find any periodic signal with shorter period at higher energy channel. In other words, QPPs with shorter period are only observed at lower energy channels, but fail to be detected at higher energy channel.
# Discussions
It is very interesting that the flare QPPs display multiple periods in HXR emission during the impulsive phase of solar flare, and the shorter periods tend to appear at lower energies of the X-ray photons, while the longer periods appear at higher energies, with some intersection of the periods at medium energies. The periods are 27 s and 37 s at lower photon energies (Figures [\[qpp1\]](#qpp1){reference-type="ref" reference="qpp1"} and [\[qpp2\]](#qpp2){reference-type="ref" reference="qpp2"}), and become around 80 s at higher photon energies (Figure [\[qpp4\]](#qpp4){reference-type="ref" reference="qpp4"}). Then the periods of 27 s, 46 s and 60 s are found at medium photon energies (Figure [\[qpp3\]](#qpp3){reference-type="ref" reference="qpp3"}). To our knowledge, this is the first report that the QPPs display their dominant periods become longer with the increasing of the photon energies in HXR emission. Previous observations have reported the various periods in multi-wavelengths. For example, report 16 s and 36 s QPPs at SXR and HXR channels (*RHESSI*, 3\(-\)`<!-- -->`{=html}50 keV). detect flare QPPs with periods of 24\(-\)`<!-- -->`{=html}160 s at EUV band (AIA 171 Å). investigate QPPs during the decay phase of solar flare and find double periods of \(\sim\)`<!-- -->`{=html}53 s and \(\sim\)`<!-- -->`{=html}72 s at SXR channels (*RHESSI*, 3\(-\)`<!-- -->`{=html}25 keV). However, these observations just report the multiple periods at one or several bands, but not indicate the periods changing with the photon energies. Another interesting thing is that the flare QPPs show their multiple periods appear at different times during the impulsive phase of solar flare. This is also a new result comparing with previous findings about multi-period QPPs. For example, find that the periods become longer continuously in the nonthermal emission. and detect the different periods during the flare impulsive and decay phases, respectively. It is also interesting that the flare QPPs show the period ratio within several values at HXR channels. At lower photon energies, the period ratio is estimated to about 1.37 from the two dominant periods; and at medium photon energy, the period ratios are around 2.22, 1.70 and 1.30, while at higher photon energy, there is only one period value. On the other hand, the period ratio between different photon energies can also be calculated. For example, the period ratios are 2.96 and 2.16 between higher and lower photon energies, and the period ratios of around 2.96, 1.74 and 1.33 can be detected between higher and medium photon energies. The period ratios between medium and lower photon energies are much more and complex, since both of them display several dominant periods. Similar as previous findings, all these period ratios are deviating from two. This may be due to the expansion of flare loop, or the density stratification in solar flare. The X1.6 flare on 2014 September 10 also exhibits QPPs with multiple periods at various wavelengths, such as \(\sim\)`<!-- -->`{=html}240 s QPPs at HXR, EUV/UV and radio emission, \(\sim\)`<!-- -->`{=html}120 s and \(\sim\)`<!-- -->`{=html}60 s QPPs at SXR/HXR and EUV/UV bands. In this paper, we obtain the multiple periods in HXR emission, i.e., 27 s, 37 s, 46 s, 60 s and 80 s. The different periods of the flare QPPs in pervious findings are derived from the different rapidly varying signals, which are dependent on the window timescale. The window timescale of \(\sim\)`<!-- -->`{=html}240 s QPPs is 256 s, and that of \(\sim\)`<!-- -->`{=html}60 s QPPs is 100 s. Both these QPPs can last up to around 18:00 UT at EUV/UV bands, which is the decay phase of solar flare. Different from these previous observations, we find the shorter and then the longer periods during the impulsive phase of this flare. It is well known that the X-ray emission at lower and higher photon energies can be produced by different mechanisms, i.e., by thermal plasmas or by nonthermal electron beams. Usually, the thermal component is extended up to 20 keV. And sometimes, the nonthermal component can also be as low as about 10 keV. Therefore, the longer periods at medium and higher photon energies are produced by accelerated particles. And the shorter periods at lower photon energies (12.0\(-\)`<!-- -->`{=html}27.3 keV) seem to be produced by thermal plasmas. However, we could not find the shorter periods at 4.6\(-\)`<!-- -->`{=html}12.0 keV, and the shorter periods can also be detected at 27.3\(-\)`<!-- -->`{=html}50.9 keV. This is an indirect evidence that the QPPs are the features of nonthermal emission. In a word, all the periods of flare QPPs in X-ray emission during the impulsive phase of solar flare may be produced by nonthermal electron beams. Meanwhile, fitting the energy spectrum could proof this hypothesis, but it is out of scope of this paper. | {'timestamp': '2017-06-07T02:04:59', 'yymm': '1706', 'arxiv_id': '1706.01680', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01680'} |
# Introduction {#sec:intro}
Conversational agents can seamlessly integrate into our lives by offering a natural language interface for complex tasks. However, the complexity of conversations with current slot-filling dialogue systems is limited. One limitation is that the user usually cannot refer back to an earlier state in the dialogue, which is essential *e.g.*, when comparing alternatives or researching a complex subject. The recently published Frames dataset provides 1369 goal-oriented human-human dialogues where the participants had to decide on a vacation package to purchase. The authors observed that in order to make up their minds, participants often compared different packages and referred to items that had been previously discussed during the dialogue. Current dialogue systems do not model the dialogue history in a way that a user can go back-and-forth between the different things that have been discussed. To address this shortcoming, introduced a new task called *frame tracking*. Frame tracking is an extension of the state tracking task. In a task-oriented dialogue system, the state tracker keeps track of the user goal. The user goal is often represented as the set of constraints that the user has (*e.g.*, a budget) as well as the questions that the user has about the items presented to her by the dialogue system (*e.g.*, the price of the vacation package). It is assumed that the dialogue system only needs to keep track of the last set of constraints given by the user. As a consequence, the user can change her goal during the dialogue but never come back to a previous goal. Frame tracking consists of recording all the different goals set by the user during the dialogue. This requires creating a new frame for each new user goal, which is the annotation provided with the Frames corpus. A frame tracker needs to be able to assign each new user utterance to the frames it references. This requires understanding which frame the user is talking about and recognizing when the user changes her goal, which implies that a new frame is created. For *e.g.* comparisons, multiple referenced frames need to be identified. This paper proposes a neural model that attempts to solve these tasks. We show that the model significantly outperforms the baseline proposed by on all the tasks required to perform frame tracking except for when the user switches frames without specifying slots. We also provide an analysis of frame tracking. In particular, we show that our model knows what frame anaphora refer to almost 90% of the time, and which hotel is being talked about 84.6% of the time. On the other hand, it does not perform well on slots which tend to be repeated in many frames, such as `dst_city` (destination city). It also has difficulties selecting the right frame among similar `offer`s introduced in the same dialogue turn.
# Frame Tracking: An Extension of State Tracking {#sec:frametracking}
In a goal-oriented dialogue system, the state tracker records the user goal in a semantic frame. The Dialogue State Tracking Challenge (DSTC) defines this semantic frame with the following components:
- User constraints: slots which have been set to a particular value by the user.
- User requests: slots whose values the user wants to know.
- User's search method: the user's way of searching the database (*e.g.*, by constraints or alternatives).
In state tracking, when a new user constraint is set, it overwrites the previous one in the frame. In frame tracking, a new user constraint creates a new frame and thus, there are as many frames as user goals explored during the dialogue. To deal with user goals, two components specific to this setting were added to a frame, namely:
- User's comparison requests: user requests for this frame and one or more other frames.
- User's binary questions: user questions with slot types and slot values for this frame and possibly one or more other frames.
The user's search method is not part of the semantic frame defined for frame tracking. In addition, a frame is also created when the wizard proposes a vacation package to the user. This type of frame contains the same components as the ones defining a user goal except that the constraints are set by the wizard and not the user. A new frame is created so that if the user wants to consider the package, it is possible to switch to this frame and consider it to be the current user goal. An example of comparison request is "*Could you tell me which of these resorts offers free wifi?*" and an example of binary question is "*Is this hotel in the downtown area of the city?*" or "*Is the this trip cheaper than the previous one?*". A user request only has a slot type, *e.g.*, "*Where is this hotel?*" whereas a binary question has a slot type and a slot value. In other words, a binary question amounts to a confirmation and a request, to an open question. Frames was collected using a Wizard-of-Oz method (WOz, [@kelley_iterative_1984; @rieser_corpus_2005; @wen_network-based_2016]): for each dialogue a user and a wizard were paired. The user had a set of constraints and was tasked with finding a good vacation package that fits these constraints. The wizard had access to a database of vacation packages and helped the user find a suitable package. The wizards were thus playing the role of the dialogue system. Each dialogue turn in the dataset is annotated with the currently active frame, *i.e.*, the frame that is being currently discussed. The corpus was annotated in such a way that both users and wizards could create new frames. On the wizard side, a new frame is created whenever the wizard proposes a new package to the user. However, only the user can switch the currently active frame, for instance, by asking questions about a package proposed by the wizard. The motivation is that the user should have control of the frame being discussed throughout the dialogue, the dialogue system being an assistant to the user. define the frame tracking task as follows:
This task is illustrated in [\[fig:frametracking\]](#fig:frametracking){reference-type="ref" reference="fig:frametracking"}. We propose a model that tries to solve this task and analyze this model's behavior on several sub-tasks.
# Related Work
As discussed in the previous section, frame tracking extends state tracking from only tracking the current user goal to tracking all the user goals that occur during the dialogue. Recently, several approaches to state tracking have attempted to model more complex behaviors than sequential slot-filling. The closest approach to ours is the Task Lineage-based Dialog State Tracking (TL-DST) setting proposed by. TL-DST is a framework that allows keeping track of multiple tasks across different domains. Similarly to frame tracking, propose to learn a dynamic structure of the dialogue composed of several frames corresponding to different tasks. TL-DST encompasses several sub-tasks among which *task frame parsing* which consists of assigning a set of new dialogue acts to frames. This relates to frame tracking except that they impose constraints on how a dialogue act can be assigned to a frame and a dialogue act can only reference one frame. trained their tracking model on datasets released for DSTC (DSTC2 and DSTC3, [@Henderson:14; @Henderson:14c]) because no appropriate data for the task was available at the time. With this data, they could artificially mix different tasks within one dialogue, *e.g.*, looking for a restaurant and looking for a pub, but they could not study human behavior and how humans switch between tasks and frames. Besides, TL-DST allows switching between different tasks but does not allow comparisons which is an important aspect of frame tracking. Another related approach was proposed by, who re-interpreted the state tracking task as a question-answering task. Their state tracker is based on a memory network and can answer questions about the user goal at the end of the dialogue. They also propose adding other skills such as keeping a list of the constraints expressed by the user during the dialogue. This work did not attempt to formalize the different constraints as separate states to record. Before describing our frame tracking model, we analyze the frame-switching and frame-creation behavior in Frames.
# Analysis of Frame References {#sec:frames-analysis}
## Reasons for Referencing Other Frames {#sec:frames-reasons}
The Frames dataset contains 19986 turns, among which 10407 are user turns. In 3785 (36%) of these user turns, the active frame is changed. When the active frame is not changed, the user refers to one or more other frames in 7.5% of the turns. If we consider only `inform` acts[^1], a dialogue system with a traditional state tracker which tracks only a single semantic frame would be able to deal with the subset of frame changes which correspond to overriding an already established value (1684 turns or 44% of the turns where frame changes occur). The remaining 2102 (56%) turns contain `switch_frame` acts from the user. The `switch_frame` act indicates when a user switches from the currently active frame to a previously-defined frame. A `switch_frame` act directly follows one or several vacation-package offers from the wizard in 1315 (38%) of the frame-changing turns. In 428 turns, the user selects between multiple offers made by the wizard and in 887 turns, she accepts a single offer made by the wizard. A total of 787 (20%) frame switches are made to a point in the dialogue which is anterior to the directly preceding turn. Note that in this work, we assume that we know the list of all previous frames at each turn of the dialogue but a practical dialogue system should generate this list dynamically during the dialogue. For this reason, it is crucial to also correctly interpret the user's `inform` acts so that if a user appeals to an old frame, this frame exists and is correctly identified. Most of the turns where the user does not change the active frame but refers to other frames contain `request_compare` (asking to compare different frames, 191), `negate` (98), `request` (28), and `request_alts` (asking for another package, 17) acts.
## Examples {#sec:examples}
In this section, we categorize instances of interesting frame-related user behavior and discuss the resulting requirements for a frame tracker.
- **Switching to a frame by mentioning a slot value.** "Oh, the Rome deal sounds much better!", "Can you tell me more about the Frankfurt package?", "I'll take the 13 day trip then!". For this case, we need to find which frames match the identified slot values, for instance, the destination city in the first example. Since there might be multiple matching frames, we have to incorporate recency information as well. In addition, equivalences have to be taken into account (13--thirteen, September--sept, NY--Big Apple, etc.). Furthermore, in some cases, we need to learn equivalences between slots. *E.g.*, the user has a budget, but the wizard typically only mentions prices.
- **Switching to a frame without referencing it directly, usually by accepting an offer explicitly or implicitly.** "yeah tell me more!", "yes please", "Reasonable. any free wifi for the kids?". The difficult part here is to identify whether the user actually accepted an offer at all, which also modifies the frame if the user asks follow-up questions in the same turn like in the third example. Some users ignore irrelevant wizard offers completely.
- **Switching to a frame using anaphora.** "Yeah, how much does the second trip cost?", "When is this trip and what is the price?", "Give me the first option, thank you". This is a slightly more explicit version of the previous case, and requires additional logic to determine the referenced frame based on recency and other mentioned slot values.
- **Implicit reference for comparisons.** "Do these packages have different departure dates?".
- **Explicit reference for comparisons.** "Can you compare the price of this and the one to the package in St. Luis?" (sic)
- **Creating a new frame by specifying a conflicting slot value.** "okaaay, how about to Tijuana then?", "what's the cheapest you got?", "Can I get a longer package if I opt for economy first?" Here, the mentioned slot values need to be explicitly compared with the ones in the current frame to identify contradicting values. The same similarities discussed in frame switching above must be considered. The context in which the slot values occur may be crucial to decide whether this is a switch to an old frame or the creation of a new one.
- **Creating a new frame with an explicit reference to a previous one.** "Are there flights from Vancouver leaving around the same time from another departure city?", "I'd like to also compare the prices for a trip to Kobe between the same dates.", "Is there a shorter trip to NY?". In these examples, the slots time, date, and duration depend on references to frames (the current frame and the NY frame, respectively).
# Frame Tracking Model {#sec:methods}
In the previous section, we identified various ways employed by the users to reference past frames or create new ones. In the following sections, we describe a model for frame tracking, *i.e.*, a model which takes as input the history of past frames as well as the current user utterance and the associated dialogue acts, and which outputs the frames references for each dialogue act.
## Input Encoding {#sec:methods-in-enc}
Our model receives three kinds of inputs: the frames that were created before the current turn, the current turn's user dialogue acts without frame references, and the user's utterance. We encode these three inputs before passing them to the network. The frames and the dialogue acts in particular are complex data structures whose encodings are crucial for the model's performance.
### Text Encoding {#sec:methods-text-enc}
We encode the user text as well as all the slot values by tokenizing the strings[^2] and converting each token to letter trigrams[^3]. Each trigram \(t\in\mathcal T\) is represented as its index in a trainable trigram dictionary \(D_{\mathcal T}\).
### Frame Encoding {#sec:methods-frame-enc}
We encode only the constraints stored in the set \(\mathcal F\) containing the frames created before the current turn. In the Frames dataset, each frame \(F\in\mathcal F\) contains constraints composed of slot-value pairs, where for one slot \(s\in\mathcal S\) multiple equivalent values (*e.g.*, NY and New York) and additional negated values (for instance if the user says that she does not want to go to a city proposed by the wizard) may be present. We encode a string representation of the most recent non-negated value \(v\) as described in [5.1.1](#sec:methods-text-enc){reference-type="ref" reference="sec:methods-text-enc"}. The slot type is encoded as an index in a slot type dictionary \(D_{\mathcal S}\). The final frame encoding is the concatenation of all slot-value pairs in the frame. In addition to the encoded frames, we also provide two vectors to the model: a one-hot code \(f_{c}\) marking the frame that was active in the last turn (the bold frame in [\[fig:frametracking\]](#fig:frametracking){reference-type="ref" reference="fig:frametracking"}) and a one-hot code \(f_{n}\) marking the frame that will be added if a new frame is created by the user in this turn (the frame marked "(new)" in [\[fig:frametracking\]](#fig:frametracking){reference-type="ref" reference="fig:frametracking"}).
### Similarity Encoding {#sec:methods-similarity-enc}
To simplify learning of plain value matching, we precompute a matrix \(S_L\in \mathbb{R}^{N{\times}\mathcal F}\), which contains the normalized string edit distance of the slot values in the user act to the value of the same slot in each frame, if present.
### Recency Encoding
We also provide the model with information about the history of the dialogue by marking recently added as well as recently active frames, coded as \(h^\tau_{d}\) and \(h^\tau_{c}\), respectively, at turn \(\tau\). For a frame \(f\) introduced or last active at turn \(\tau_f\), we set \[\begin{aligned}
\label{eq:recent_discount} h_\cdot^\tau(f) = \begin{cases} 0 & \text{~if~} \tau < \tau_f\\ 1 & \text{~if~} \tau = \tau_f\\ \gamma h_\cdot^{\tau-1} & \textnormal{otherwise.}\\ \end{cases} \end{aligned}\]
### Act Encoding {#sec:methods-act-enc}
A dialogue act in the current turn has an act name \(a\in\mathcal A\) and a number of arguments. Each argument has a slot type \(s\in\mathcal S\) and an optional slot value \(v\). We use a dictionary \(D_{\mathcal A}\) to assign a unique index to each act \(a\), and use the same method as described in [5.1.2](#sec:methods-frame-enc){reference-type="ref" reference="sec:methods-frame-enc"} to encode slot-value pairs. In addition to the \(N\) triples \((a,s,v)\), we encode every act \(a\) separately, since an act may not have any arguments but still refer to a frame (cf. frame switching examples in [4.2](#sec:examples){reference-type="ref" reference="sec:examples"}).
## Output Encoding {#sec:methods-out-enc}
For each triple \((a,s,v)\), our model outputs a multinomial distribution \(p_{asv,F}\) over the frames \(F\in\mathcal F\). Additionally, for each act \(a\in\mathcal A\) and frame \(F\in\mathcal F\), we determine the probability \(p_{a,F}\) that \(F\) is referenced by \(a\). It can be difficult for the model to correctly predict the cases when the referenced frame is the currently active frame, especially in situations where (a) the slot values do not match and (b) the active frame was changed by an earlier act within the same turn. To address this challenge, in the target, we replace all occurrences of the active frame with a special frame with index 0. In the example of [\[fig:frametracking\]](#fig:frametracking){reference-type="ref" reference="fig:frametracking"}, the value `flex=T` would point to this frame 0 since the active frame is changed by a previous value, in this case, the `budget`. In the loss function, we do not penalize the model for confusing the active frame and the special frame except for `switch_frame` and frame-creating `inform` acts, for which we want the model to predict the referenced frame. During prediction, we distribute \(p_{asv,0}\) over \(\mathcal F\) according to the predicted active frame: \[\begin{aligned}
g_s & = \begin{cases} 1 & \text{if a \texttt{switch\_frame} act is present}\\ 0 & \text{otherwise} \end{cases}\\ p_{\text{new}} & = p_{asv,|\mathcal F|+1}\\ p_{asv,F} &:= p_{asv,F} + p_{asv,0} \big( (1-g_s) \times p_{\text{new}} \\ & +g_s \times p_{\text{switch},F}\big), \end{aligned}\] where \(p_{\text{switch},\cdot}\) is the distribution assigned by the model to the `switch_frame` act. If no new frame was predicted and no `switch_frame` act is present, the remaining probability mass is assigned to the previously active frame.
We further analyze the difference in performance between our frame tracking model and the rule-based baseline on classes of predictions on a single fold of the data. We organize the turns in the test set into 11 classes and measure performance by computing accuracy only on turns that fall into the respective class. We first observe that the baseline model almost completely fails to identify frame changes when a new value is introduced by a user (4% accuracy over 303 turns), frame changes associated with `switch_frame` acts that do not have slot values, or when a `switch_frame` act is present in a turn following one that does not contain an `offer` act. On the other hand, the baseline model predicts lacks of frame changes (74.3% over 1111 instances) and frame changes after an `offer` (62.2% over 312 instances) quite well. Our model dominates the rule-based baseline on all classes except for the prediction of frame changes with `switch_frame` acts that do not have slot values (4.2% over 24 occurrences). Partial comparison results are presented in [\[tbl:event_types\]](#tbl:event_types){reference-type="ref" reference="tbl:event_types"}. Perhaps surprisingly, our model correctly predicts 70.5% of frames associated with `request_compare` acts whereas the baseline only correctly identifies 40.9% of them. We then computed the accuracy on the set of unique act and slot combinations in the dataset. Here, our model outperforms the baseline on all act-slot pairs with more than 10 occurrences in the test set. We observe that the baseline performs quite poorly on `switch_frame`s with `dst_city` (destination city) slots, whereas our model does not have such a drawback. The same is true for a `switch_frame` with a `duration` or for an `inform` with a `seat` (economy or business flight seat) or even a `request` with an `end_date`. Results are presented in [\[tbl:act_slot\]](#tbl:act_slot){reference-type="ref" reference="tbl:act_slot"}. We note that our model performs worse on combinations that should express a match with a frame whose slot values use very different spellings (such as rich abbreviations and synonyms) whereas the baseline model is the weakest when slot values can be easily confused for values of other slots (*e.g.* a rating of 5 (stars) vs. a duration of 5 (days)). Our model is also currently unable to distinguish between similar offers introduced in the same turn. Code to generate the full set of metrics will be made available.
## Lesion Studies
To assess which of the features are useful for the model, we remove the model's inputs one at a time and measure the model's performance. Results are shown in [\[tbl:lesion\]](#tbl:lesion){reference-type="ref" reference="tbl:lesion"}. We observe that the model stops learning (*i.e.* its performance does not exceed the baseline's) on the act-slot-value triples when any of the input is removed except for the new frame history, new frame candidate, and previous frame inputs. Similarly, the model performance suffers when the new frame candidate, any historical data, or the frames are removed. We observe that all the inputs are used by the model in its predictions either for \(p_{asv,F}\) or \(p_{a},F\).
# Discussion
Our model makes use of the text to correctly predict the frames associated with acts. Dependence on input text means our method is domain-dependent. The annotation process for the Frames dataset is costly, so it would be beneficial if we could transfer learned frame switching behavior to other domains, possibly with already existing NLU components. A possible solution might be to standardize the text after NLU, and use anonymous placeholders instead of domain-specific words. Additionally, our current model assumes a perfect NLU to provide acts, slots, and values as inputs. While this is helpful for researching the frame referencing issues in isolation, both components should work together. For example, currently, we assume that a `switch_frame` act is correctly identified, but we do not know the frame the user wants to switch to. In a more realistic pipeline, these decisions are closely related and also need to take more of the dialogue history into account.
## Conclusion
In this paper, we provided a thorough analysis of user behavior concerning switching between different user goals in the Frames dataset. Based on this analysis, we have designed a frame tracking model that outperforms the baseline of by almost 20% relative performance. This model assigns the dialogue acts of a new user utterance to the semantic frames created during the dialogue, each frame corresponding to a goal. We analyzed the strengths and weaknesses of the rule-based baseline and of our model on different subtasks of frame tracking. Our model outperforms the baseline on all but one subtasks. We showed that further improvement is necessary for matching slot values when they are present in many distinct frames. We have demonstrated that the frame tracking task can be performed effectively by learning from data (our model correctly identifies frame changes in about 3 out of 4 cases). This represents a first step toward memory-enhanced dialogue systems which understand when a user refers to an older topic in a conversation and which provide more accurate advice by understanding the full context of a request.
[^1]: Utterances where the user informs the wizard of constraints.
[^2]: using nltk's `TweetTokenizer`, [www.nltk.org](www.nltk.org)
[^3]: *E.g.*, "hello" is converted to \#he, hel, ell, llo, lo\# | {'timestamp': '2017-06-07T02:05:10', 'yymm': '1706', 'arxiv_id': '1706.01690', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01690'} |
null | null |
# Introduction
Heusler alloys, a vast family of ternary compounds, are often considered an ideal platform for engineering and designing novel functional materials. Such class includes both metals and insulators, and among them superconductors, topological insulators, thermoelectric alloys, and both optical and magnetic materials . As such, the possibility of using alloys of this family for fine tuning and controlling the electronic structure and the magnetic order is tantalizing. However, despite several decades of intense investigation and accumulated understanding on the Heuslers compounds , the tuning of their properties still proceeds via chemical intuition in a slow trial-and-error mode. It is then an intriguing prospect to explore more high-throughput methods for materials screening and understand whether these can identify novel designing rules. Reliable and low-cost computational methods now allow one to perform systematic investigations of large regions of the chemical space. This is known as the computational high-throughput approach . The analysis of the generated data has lead data-mining and machine-learning techniques to become part of the material science toolbox (for a non-exhaustive list of currently available materials databases see \[\]). Materials can be classified by using descriptors , simple proxies for sometime complex materials characteristics, and system properties estimated via machine leaning regression and classification . The latter are particularly useful when a direct calculation is prohibitive. An approach based on the machine learning uses statistical inference for predicting properties of a given system without performing an actual electronic structure calculation. This enables a fast, objective and cost-effective analysis of large amounts of multi-dimensional data, making machine learning a natural extension of the computational high-throughput strategy. Here we use machine learning (ML) techniques to study the magnetism of Fe-containing Heusler alloys. Iron offers a large magnetic moment, second only to Mn, but in contrast to Mn that is known to maintain a high-spin state in Heusler compounds , Fe is more susceptible to changes in the local chemical environment. This makes it an ideal choice for exploring the predictive power of ML techniques, when applied to magnetic materials. In this work we show that knowing the composition of the first two Fe coordination shells is sufficient to accurately estimate its magnetic moment using ML regression. By combining ML and density functional theory (DFT) data we are able to identify and explain trends in the compounds magnetic moment. Finally, the ideal prototype for an Fe-based regular Heusler alloy is proposed. The ML regression is used to rapidly characterise and rank all possible alloys of the proposed prototype, according to the maximal attainable magnetization. We note that all predictions are made by only using a list of Fe neighbours, without any need for additional *ab initio* calculations. This demonstrates the potential of the machine learning approach for developing a fast, high-volume, method for screening magnetic materials. The paper is organized as follows. In the next section we describe the methods at the foundation of the machine-learning process and the general attributes that enter into the description of magnetism in Heusler alloys. Then we introduce our results focussing on the role of the nearest neighbour and of next nearest neighbour coordination, and defining the physical origin of the magnetic moment trends. Then we use our machine learning scheme to identify magnets with large magnetization. Finally we conclude.
# Method
## General Considerations
Structure-to-property relations are at the heart of all problems in the material science. These are implicitly determined by the electronic structure of any given compound, which nowadays is routinely computed by using *ab initio* methods. Accurate information about various material properties can then be extracted solely from theory. Machine learning (ML) allows us to take a rational approach to large-scale material investigation. The underlying assumption is that once there is enough materials data available, an answer to the structure-to-property question should be already implicitly contained in the data. We can thus speak about "learning from the data". The ML methods are built specifically for this task, providing us with a practical mean to construct approximate structure-to-property relationship maps with a well defined domain of validity. The latter, however, needs to be established through tests. A major advantage of the ML approach is that it thrives on large datasets, offering a high-throughput, objective analysis of the material properties. The data are never discarded, but instead they are continuously integrated to refine the predictions and can be reused to address new questions. Thus, it is an inductive, data-driven, approach of performing material research. The trade-off is that the ML results are usually less accurate than those obtained by using *ab initio* methods. Here we focus on "supervised learning" methods, which include regression and classification algorithms. The advantage of such class of schemes is that the quality of the ML predictions can be evaluated, for instance in the case of the regression by calculating a mean-square error. The purpose of the training procedure is then that of minimizing the risk of making incorrect predictions. The idea is that a well-trained algorithm may produce a significant error for an individual system but it shall perform in a satisfactory manner for the entire data set. An outline of the general procedure used in this work is shown in figure [\[fig_ml_process\]](#fig_ml_process){reference-type="ref" reference="fig_ml_process"}. The process starts by defining the problem and by selecting suitable data to describe it. The data may come from different sources and is usually combined into suitable input features. For example, in figure [\[fig_cluster\]](#fig_cluster){reference-type="ref" reference="fig_cluster"} we use a cluster of atoms to define the magnetic moment of the central atom and various atomic data (see equation [\[eq_amag_features\]](#eq_amag_features){reference-type="ref" reference="eq_amag_features"}) are considered to describe the cluster further. Next, the available data are split into the training and the test dataset, according to the output property that one wishes to evaluate. Here it is crucial to preserve the underlying property distribution when the data is split, otherwise one may end up with a biased training set. If the training set is biased, the ML model may not be predictive for data outside it, and furthermore the fidelity of the algorithm may be erroneously estimated. In brief the ML model trained on such a dataset will usually perform badly on new data, i.e. it will not be predictive. The test step is used as an independent check of the ML accuracy and its ability to generalize to new data. The test dataset is never used for building the ML model. The *model building* phase, depicted as a single step in Fig. [\[fig_ml_process\]](#fig_ml_process){reference-type="ref" reference="fig_ml_process"}, is actually an iterative two-step procedure. First a set of input features, namely an input vector encoding a number of chemical/physical properties, is constructed from the raw input data and then different ML algorithms are trained using that input. The latter step includes the choice of the, e.g., regression algorithm, and the optimization of the hyperparameters. These two steps are repeated until a satisfactory accuracy is achieved. The choice of the input data and its transformation into a useful set of input features is the most important aspect of the entire process. In this step we use our domain knowledge to convert the raw data into descriptive features, which correlate with the output. This describes an inductive approach for constructing a ML model, which is feasible when we have some understanding of the underlying processes correlating the input to the output. In material science this should often be the case. A minimal number of input features, which are included in the model, leaves a possibility of its interpretation. Alternatively, one needs to follow a deductive approach. In this case one starts from a as-large-as possible number of input features and performs an input reduction analysis, eliminating the variables that do not correlate with the output. There is usually no a unique way to choose the input features and interpret the ML model. If the dimension of the input space is not large one may seek to explore the significance of the individual input features. This can help in deepening the knowledge of the system under investigation and guide further *ab initio* calculations. However, such step is not always possible especially when the dimension of the input space is large. The importance of having a working ML description of the system is not diminished by this feature. It is often much more practical to explore the data using a ML algorithm than working with the raw data. For example, in the case of a ML regression, one deals with a single function (a map), instead of a large database. Exploring the connection between the variables is thus much simpler and faster. The added benefit is that the ML algorithm will often accurately interpolate where data is missing.
## Machine Learning Model for Magnetism
In this work we use data extracted from an in-house-made Heusler alloy database, named *Materials Mine* . All calculations have been performed using the PAW pseudopotential implementation of DFT contained in the VASP code and the generalized gradient approximation of the exchange and correlation functional as parametrized by Perdew, Burke and Ernzerhof . For each chemical composition, the ground state is calculated for different site occupations, structural parameters and sublattice magnetic order. Furthermore, we complement the DFT data by various atomic properties information obtained from a wide range of sources in literature, including both experimental and theoretical data. The sources will be properly cited individually whenever used. The knowledge of the crystal cell volume, namely the inter-atomic distance, is vital for studying electronic structure properties and magnetism in particular. It is reasonable to expect that this quantity will repeatedly appear in all ML models and we wish to be able to predict it without relying on the DFT data. We have then trained a volume regressor using the DFT data for the 229 fully relaxed full-Heusler structures, having the lowest energy for a given composition and site occupation. We note that the number of different compositions in the database is larger than 229, but we restrict out choice to this dataset since the same was used with success for other investigations. In any case we will show that this choice does not affect the quality of the final result. The ML model was built using the ridge regressor algorithm as implemented in Scikit-learn package . We have used 30 % of the dataset for the test and the rest was used as training set. The input vector was constructed by including the atomic numbers, the atomic volumes, and the atomic radii of the three nonequivalent ions. The atomic volumes were obtained from Mentel and then the atomic radius was calculated for each element. Here we assume atoms to be homogeneous solid spheres. A root mean square (RMS) error on the test data was calculated to be 3.16 Å^3^. In comparison, the mean volume of alloys in the dataset was \(\approx\) 63 Å^3^. The attained accuracy of the ML algorithm is thus comparable to the precision of the DFT calculations, namely is a good predictor for the volume. We now wish to estimate the magnetic moment of Fe-containing Heusler alloys by using the ML approach. The magnetic moment of 3\(d\) transition metals is well localized and can be understood as an atomic moment, which gets modified by its local surrounding. It is therefore interesting to try to relate the environment of an atom to its moment. The DFT data was used to construct clusters of atoms representing the local environment of the central atom, as shown in figure [\[fig_cluster\]](#fig_cluster){reference-type="ref" reference="fig_cluster"}. We construct one cluster for each atomic site of the parent alloy, namely we construct 4 clusters per DFT calculation (per Heusler prototype). The data corresponding to the lowest energy states having a given formula unit was selected from the database. Note that here we consider a much larger, and less constrained, set of calculations than before. In fact, we construct 18,268 clusters, of which \(\approx\) 7,000 were used for the test. The site projected magnetic moment of the central atom, obtained from the DFT results, was used as the target property. The input vector was constructed as \[\label{eq_amag_features} \vec{v}_\mathrm{in}=\left(\{Z_i\}, R_0, a_{\mathrm{lat}}, \{r_{0i}\}, \{N_i\}, S_0 \right)\:,\] where \(\{Z_i\}\) (\(i=0,1,2,3\)) and \(R_i\) are the atomic number and the atomic radius of the \(i\)-th atom, respectively, and \(N_i\) is the valence (\(i=0,1,2\)). The atomic positions are labeled as in Fig. [\[fig_cluster\]](#fig_cluster){reference-type="ref" reference="fig_cluster"}. Here \(r_{0i}\) is the distance between 0-th and the \(i\)-th atom, scaled by the sum of their atomic radii. The "effective" cubic lattice constant, \(a_{\mathrm{lat}}\), is calculated from the volume of the parent Heusler structure, which in turn is estimated using the previously discussed regression model. Finally, \(S_0\) is the Stoner parameter of the central atom, obtained from Janak . As for as the regression is concerned, we have found the Random Forest Regression, as implemented in the Scikit-learn library , to give the best results. The RMS error measured on the test dataset is 0.4, with the RMS error on the subset having Fe as central atom being measured somewhat higher, \(\sim\) 0.54. We did not notice any improvement in the RMS error when we performed the training by only using Fe-centered clusters. We believe there are two reasons for this finding. The first is the reduced dataset size utilized. We find that \(\sim\) 7000 clusters are barely sufficient to converge the algorithm learning curve. The second is the that the larger dataset also contains non-magnetic atoms, whose magnetic moment is trivial to estimate. We note that in some cases the errors are much larger than the RMS value, which seems to be at least partly related to the observed convergence issues in the high-throughput calculations. For example, the DFT data is found to exhibit a large variation in the magnetic moment of Fe (see Fig. [\[fig_effect_of_z3\]](#fig_effect_of_z3){reference-type="ref" reference="fig_effect_of_z3"}). Reliable methods for data cleaning are needed, however, in this work we treat such anomalous calculations simply as a noise.
# Results
## Role of the Coordination Shells
In the Methods section we have presented a ML model for predicting the magnetic moment of an atom embedded in a Fe-containing regular Heusler alloy (see figure [\[fig_cluster\]](#fig_cluster){reference-type="ref" reference="fig_cluster"}). The model estimates the magnetic moment based on four key variables: the three atomic numbers specifying the coordination of the central atom, \(\{Z_i\}\), and the lattice constant of the parent Heusler alloy. The latter can be estimated by using a ML regression (see Table [1](#tbl_estimates){reference-type="ref" reference="tbl_estimates"}). The atomic properties, needed to construct the input vector (Eq. [\[eq_amag_features\]](#eq_amag_features){reference-type="ref" reference="eq_amag_features"}), are easily obtained. This makes the method completely free of input *ab initio* parameters. We then use this model to explore and gain a deeper understanding of our DFT data. Here we focus on \(X_2\)Fe\(Z\) alloys, where the central atom is Fe, with the cubic L2\(_1\) crystal symmetry and the corresponding tetragonal structures . In this case the two inequivalent sites in the first coordination shell are occupied by identical atoms (i.e. \({Z}_1 ={Z}_2 = {X}\) in Fig. [\[fig_cluster\]](#fig_cluster){reference-type="ref" reference="fig_cluster"}), a fact that allows us to explore the effects of three structural parameters.
### Role of the Next-Nearest Neighbour
We first look at the composition of the second coordination shell, i.e. the effect of \(Z_3\) of the magnetic moment of Fe, \(m_\mathrm{Fe}\) (values are provided in unit of Bohr magneton, \(\mu_\mathrm{B}\)). The DFT dataset was sampled at a constant volume, namely the Wigner-Seitz radius was set to \(R_\mathrm{WS}=2.7 \mathrm{\si{\bohr}}\), and the 8 nearest neighbour atoms were fixed to Fe. This Wigner-Seitz radius roughly corresponds to a Heusler lattice constant of 5.8 Å. The corresponding data and the ML estimate of the magnetic moment are shown in Fig. [\[fig_effect_of_z3\]](#fig_effect_of_z3){reference-type="ref" reference="fig_effect_of_z3"}(a). Visual inspection of the data reveals that transition metals (\(21<Z_3<30\), \(39<Z_3<48\)) and main group elements (\(13<Z_3<16\), \(31<Z_3<34\), \(49<Z_3<52\)) make two distinct classes of next-nearest-neighbours. On the one hand, transition metals tend to increase the magnetic moment of Fe proportionally to their valence. On the other hand, main group elements tend to cluster and yield a maximal magnetic moment. In particular, for a given nearest neighbour and volume \(m_\mathrm{Fe}\) is only weakly affected by the choice of the main group element at the \(Z_3\) site. This can be seen, for example, in figure [\[fig_fe_tm_all_volume_dep\]](#fig_fe_tm_all_volume_dep){reference-type="ref" reference="fig_fe_tm_all_volume_dep"}, where the magnetic moment of Fe appears not to be correlated to the atomic number of the next-nearest neighbour, \(Z_3\). Here we would like to point out that this does not contradict the well established Slater-Pauling rule . In particular, for Co\(_2XY\) alloys this rule would imply that the net cell moment, \(m_\mathrm{cell}\), scales linearly with the valence of the alloy, \(N_\mathrm{V}\), and reaches maximum when \(N_\mathrm{V}=30\). As an example, for Co\(_2\)FeAl, Co\(_2\)FeSi and Co\(_2\)FeP alloys (\(N_\mathrm{V}=\) 29, 30 and 31) we find the DFT moments, \(m_\mathrm{cell}\), to be (per ), respectively. However, the corresponding Fe moments change only slightly, namely we find \(m_\mathrm{Fe}=\) . We note that the observed ML trend is volume dependent, giving a spurious representation of the valence trend across the transition metal series. The problem, however, does not affect the main group elements. The ML magnetization trend captured at smaller volumes, \(R_\mathrm{WS} \approx 2.4 \mathrm{\si{\bohr}}\), is qualitatively different from the trend at larger ones, \(R_\mathrm{WS} \approx 2.8 \mathrm{\si{\bohr}}\). In contrast, the DFT data trends shown in Fig. [\[fig_effect_of_z3\]](#fig_effect_of_z3){reference-type="ref" reference="fig_effect_of_z3"} remain by large unaffected by the volume change. We, therefore, find it necessary to combine the DFT data and the ML approach to obtain a complete picture. In spite of this, the numerical precision of the ML estimate is always within the limits established by the regressor test procedure, \(\sim \SI{0.5}{\bohrmagnetron}\). From a material design perspective, the learning is that the \(Z_3\) element can be chosen to ensure the stability of a Heusler alloy without compromising the magnetic moment. Figure [\[fig_effect_of_z3\]](#fig_effect_of_z3){reference-type="ref" reference="fig_effect_of_z3"}(b) shows the enthalpy of formation for the same set of Heusler alloys shown in figure [\[fig_effect_of_z3\]](#fig_effect_of_z3){reference-type="ref" reference="fig_effect_of_z3"}(a). The enthalpy of the alloy is calculated with respect to the decomposition into the most stable elemental phases, so that does not provide a strict stability criterion, but simply a guideline for stability . With a small number of exceptions we find that only the main group elements at the \(Z_3\) site have a good chance to yield thermodynamically stable alloys. Such a result could be anticipated based on the known Heusler chemistry . The freedom of the choice of \(Z_3\) opens up the possibility to tune the volume of the alloy, and to control the critical temperature . In Table [1](#tbl_estimates){reference-type="ref" reference="tbl_estimates"} we show that Heusler alloy volume can be accurately estimated using the ML regression.
### Role of the Nearest Neighbour
In the previous discussion we have shown that the magnetic moment of Fe is independent of the choice of \(Z_3\) as long as \(Z_3\) is a main group element. We now study the magnetic moment as a function of the nearest neighbour ion, \(Z_1\), keeping a main group element at the \(Z_3\) site and a fixed volume (see Fig. [\[fig_fe_tm_all_volume_dep\]](#fig_fe_tm_all_volume_dep){reference-type="ref" reference="fig_fe_tm_all_volume_dep"}). The trend with volume of \(m_\mathrm{Fe}\) depends on the choice of Z\(_1\) and it is difficult to qualify. The ML regression, however, can be used to take the volume effects into account with a good level of precision. We find that the valence of the nearest neighbour ions determines the magnetic moment of Fe for the entire range of volumes investigated. When the valence of the nearest neighbour ion is less than 8 the moment decreases, and conversely, when it is larger it increases. The same trend is found for all nearest neighbours belonging to the 3\(d\) and the 4\(d\) transition metal series, as shown in Fig. [\[fig_fe_tm_all_volume_dep\]](#fig_fe_tm_all_volume_dep){reference-type="ref" reference="fig_fe_tm_all_volume_dep"}. The maximal moment of Fe is obtained when Ni or Pd constitute the first coordination shell. The ML trend, shown with a blue line, reproduces the magnetization trend for all possible nearest neighbours. For main group neighbours the magnetization shows a strong variation with the valence, taking a minimum value in the middle of the series. The overall magnetic moment is then reduced when compared to the situation with transition metal elements. In conclusion, we have found that late transition metals: Co, Ni, Cu, Rh, Pd, and Ag, make the most desirable nearest neighbours of Fe, since they maximize its local magnetic moment.
### Physical Origins of the Magnetic Moment Trends
In order to understand the physical origin of the established trend in the transition metal series, we select a subset of the compounds shown in Fig. [\[fig_fe_tm_all_volume_dep\]](#fig_fe_tm_all_volume_dep){reference-type="ref" reference="fig_fe_tm_all_volume_dep"} for further analysis. In particular we look at the following Fe-containing alloys: Sc\(_2\)FeSi, Ti\(_2\)FeAl, V\(_2\)FeSi, Cr\(_2\)FeSi, Mn\(_2\)FeS, Fe\(_2\)FeAl, Co\(_2\)FeAl, Ni\(_2\)FeSi and Cu\(_2\)FeSi, which all possess a cubic L2\(_1\) structure. We note that, in general, the tetragonal distortion does not change the main trend, so that it is not considered here. An analysis of the site projected density of states (PDOS) reveals that both the Fe and the nearest neighbor atom remain charge neutral throughout the series. This means that Fe is always occupied by \(6\) electrons and the origin of the magnetic moment trend is solely due to the on-site charge redistribution, as shown in figure [\[fig_fe_z1_charge\]](#fig_fe_z1_charge){reference-type="ref" reference="fig_fe_z1_charge"}. By integrating the orbital resolved Fe PDOS we find that for early TMs (\(Z_\mathrm{TM} \leq 25\)) the amount of charge transferred from the minority \(t_{2\mathrm{g}}\) band to the \(e_\mathrm{g}\) spin bands is proportional to the nearest neighbour valence, with the charge of the majority \(t_{2\mathrm{g}}\) band, located deep below the Fermi level, remaining roughly the same. As a result, the Fe net magnetic moment increases proportionally to the valence of the nearest neighbour. For Mn, Fe and Co nearest neighbour we observe a reverse charge transfer, from the minority \(e_\mathrm{g}\) band to the majority \(t_{2\mathrm{g}}\) one, \(\Delta \mathrm{Q_{t_{2g}^{~\uparrow}}} \approx\) 0.4 *e*, resulting in an noticeable kink in the magnetic moment trend. Our understanding of such charge re-distribution mechanism is the following. At the begining of the transition metal series, the nearest neighbour atoms hybridize weakly with the \(t_{2\mathrm{g}}\) band of Fe, resulting in a narrow and strongly spin split Fe \(t_{2\mathrm{g}}\) band (see Fig. [\[fig_fe_z1_pdos\]](#fig_fe_z1_pdos){reference-type="ref" reference="fig_fe_z1_pdos"}). As such the energy overlap of the Fe \(t_{2\mathrm{g}}\) spin bands is initially very small. By increasing the valence of the nearest neighbour ion one leads to a stronger hybridization with the Fe \(t_{2\mathrm{g}}\) spin bands, which then get wider. Consequently, their overlap increases, giving rise to an increasingly strong repulsive Coulomb interaction. It is therefore energetically more favourable to transfer a part of the minority \(t_{2\mathrm{g}}\) charge to the \(e_\mathrm{g}\) band, which is strongly spin-split and 5 eV--6 eV wide in energy, indicating spatially delocalized orbitals. For late TM neighbours, e.g. Co (\(Z_1 = 27\)), the hybridization with Fe results in a wide majority \(d\)-band, which can now accommodate extra electrons since the Coulomb repulsion is reduced by the band broadening. At the same time the energy cost of adding extra electrons to the \(e_{\mathrm{g}}\) band is increased, as the band is nearly full. For Mn neighbours (\(Z_1 = 25\)) we find the Fe \(e_\mathrm{g}\) charge to be Q\(_\mathrm{e_g}~\approx\) 2.5 *e*. The reverse charge transfer thus reduces the Coulomb energy of the \(e_\mathrm{g}\) band. The \(d\)-band electronic structure of Fe in Sc\(_2\)FeSi, V\(_2\)FeSi and Co\(_2\)FeSi alloys, which clearly illustrate the mechanism just described, is shown in Fig. [\[fig_fe_z1_pdos\]](#fig_fe_z1_pdos){reference-type="ref" reference="fig_fe_z1_pdos"}.
## Application
### Screening of High-Magnetic-Moment Heulser alloys using ML Methods
The analysis carried out on the dependence of the Fe magnetic moment on the local chemical environment enables us to propose *LTM*\(_2\)Fe*MG* as an optimal chemical composition for a ternary Fe-based Heusler alloy with maximum \(m_\mathrm{Fe}\) (*LTM* stands for late transition metal and *MG* for main group element). We stress that our assertion applies only to L\(2_1\)-type Heusler alloys, but notably a well-established preferential site occupation rule , suggests that the proposed stoichiometry will crystallize in the required regular Heusler phase. | {'timestamp': '2017-06-07T02:08:20', 'yymm': '1706', 'arxiv_id': '1706.01840', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01840'} |
null | null |
null | null |
null | null |
# Introduction {#sec:intr}
The proton and neutron density distributions generated in the course of a heavy-ion collision, especially at a total density exceeding its normal value (\(\rho_{\rm 0}\) = 0.16 fm\(^{\rm-3}\)), are among the main points of interest of theorists, experimentalists and observers, within the communities of low and medium-energy nuclear physics and of astrophysics of compact objects. The relative asymmetry of these distributions pertains to the determination of the Equation of State (EOS) of dense matter. The EOS is usually defined by the dependence of energy of per nucleon upon density and temperature. In the limit of zero temperature, following the semi-empirical mass formula, the energy per particle can be expanded in terms of the proton-neutron asymmetry \(\delta = (\rho_{\rm n}-\rho_{\rm p})/\rho\), the local relative difference between neutron and proton densities: \[\mathcal{E}(\rho,\delta)=\mathcal{E}(\rho,\delta=0)+S(\rho) \, \delta^{\rm 2} \,. \label{eq:1}\] Here, \(\mathcal{E}(\rho,\delta=0)\) is the energy per particle of symmetric nuclear matter, \(S(\rho)\) is the symmetry energy. It can be approximated as the difference between the energy per particle for symmetric and pure neutron matter at the same total density: \[S(\rho) \equiv \frac{1}{2}\left. \frac{\partial^2\mathcal{E}(\rho,\delta)}{\partial\delta^2}\right|_{\rm \delta = 0} \approx \mathcal{E}(\rho,\delta = 1)-\mathcal{E}(\rho,\delta = 0) \,.\] As the Pauli principle influences the symmetry energy, it is common to subdivide S(\(\rho\)) into a kinetic (no interacting) Fermi-gas contribution \(S_\text{kin}\) and an interaction contribution \(S_\text{int}\), shifting the focus to \(S_\text{int}\): \[\label{eq:3} S(\rho) = S_\text{kin}(\rho) + S_\text{int}(\rho) = 12.3 \, \text{MeV} \, \Big( \frac{\rho}{\rho_0} \Big)^{2/3} + S_\text{int}(\rho) \,.\] In describing a heavy-ion collision, the main uncertainties are changes in the total density \(\rho\) = \(\rho_{\rm n} + \rho_{\rm p}\), asymmetry \(\delta\) and the magnitude and density dependence of the symmetry energy S(\(\rho\)). The properties of the symmetry energy provides an important clue as to why we have comparable numbers of protons and neutrons in nuclei and how the competition between the *pp, pn* and *nn* nucleon-nucleon interactions in nuclear environment leads to stability of nuclear systems observed in nature. No theory exists at present that would offer a fundamental understanding of the symmetry energy and its properties. For nuclei, the value of the symmetry energy at average nuclear density is reasonably well constrained by masses of heavy nuclei and other data, such as the emission of light clusters in heavy-ion collisions . However, the value and the slope of the symmetry energy, \(L = 3\rho_0 \, \partial S(\rho)/\partial (\rho)|_{\rho_0}\), and their correlation even at the normal density in symmetric nuclear matter are not known with certainty (for recent reviews cf. ). Efforts to broaden the range of densities for symmetry energy studies under laboratory conditions include pursuing observables in experiments with heavy-ion collisions and comparing them to predictions of transport models, such as the observables quantifying fragmentation and isospin diffusion in peripheral asymmetric systems . Density dependence of the symmetry energy is a necessary input to these models and is provided in various empirical forms. The validity of the assumptions is judged indirectly by comparison of the outcome of transport models with experiment. We explore the maximal densities reached separately for neutrons and protons in simulations of central heavy-ion collisions at non-relativistic and moderately relativistic energies. The asymmetry \(\delta\) in high density matter and its correlation with total density are not open to direct measurement and have to remain model dependent quantities. Depending on the type of collision simulation, either single-particle wavefunctions or phase-space distributions are dynamic quantities which are followed. Proton and neutron density profiles as a function of time of duration of an individual collision event can be simulated as a function of beam energy. The maximum proton and neutron densities provide a practical upper limit on the high-density value of \(\delta\) for each case. However, it is important to realize that in a target-projectile asymmetric collision the maximal densities and maximal asymmetry may not occur at the same place and time. In this work we consider central collisions in the systems \(^\text{40}\)Ca +\(^\text{40}\)Ca, \(^\text{40}\)Ca +\(^\text{48}\)Ca, \(^\text{48}\)Ca +\(^\text{48}\)Ca, \(^\text{100}\)Sn +\(^\text{100}\)Sn, \(^\text{100}\)Sn +\(^\text{120}\)Sn and \(^\text{120}\)Sn +\(^\text{120}\)Sn within the beam energy range of (0\(-\)`<!-- -->`{=html}800\()\, \text{MeV/nucleon}\). Calcium and tin isotopes are often chosen in experimental studies. Although the \(^\text{100}\)Sn +\(^\text{100}\)Sn system is not likely to be investigated experimentally in the near future, we include that system for completeness, as a reference. Note that the ratio of the number of neutrons to the total number of nucleons in targets and projectiles involved in the colliding systems is rather similar, being 0.5 (\(^\text{40}\)Ca and \(^\text{100}\)Sn) and 0.58 (\(^\text{48}\)Ca and \(^\text{120}\)Sn). The paper is organized as follows. The numerical methods and their physical basis are explained in Sec. [2](#sec:calmeth){reference-type="ref" reference="sec:calmeth"}. The results and discussion are presented in Sec. [3](#sec:res){reference-type="ref" reference="sec:res"}. Conclusions and outlook form the content of Sec. [4](#sec:concl){reference-type="ref" reference="sec:concl"}.
# Dynamic Models {#sec:calmeth}
Since we wish the simulation in this paper to provide benchmark values of proton-neutron asymmetry in heavy-ion collisions at low and medium beam energies, we give in this section, for benefit of a wider audience, a brief survey of the physics of the collisions and the models used, together with a summary of the parameters these models depend on. The physical basis of the models should be kept in mind when judging the results and may be helpful in finding ways for their future refinement. Heavy-ion collision dynamics changes with the incident beam energy as different physical aspects play a dominant role in the energy range of \(\sim 40\)--\(800 \, \text{MeV/nucleon}\) compared with the lower beam energies. We discuss the Boltzmann-Uhlenbeck-Uehling (pBUU) approach, used primarily for medium-to-high beam energies, in Sec. [2.1](#sec:pBUU){reference-type="ref" reference="sec:pBUU"}, and the Time-Dependent-Hartree-Fock (TDHF) method, valid only for lower beam energy collisions, in Sec. [2.2](#sec:tdhf){reference-type="ref" reference="sec:tdhf"}, reflecting different physics in the low energy region. The simulation methods employed in this work are single-particle in nature, potentially underestimating the role of fluctuations, unlike other approaches to modeling heavy-ion collisions such as Quantum Molecular Dynamics (QMD) and Antisymmetrized Dynamics (AMD). These methods use Gaussian-type packets to model the particle motion and predict observables. The nature of fluctuations at high local excitation energies is becoming similar to the fluctuations in cascade models. However, at lower energies, the nature and magnitude of fluctuations is not well established. In the present work we are interested in the maxima of statistically averaged densities reached during the compression stage of the collision. At this stage the fluctuations are short-lived and do not significantly affect the density evolution. Thus single-particle approaches such as pBUU and TDHF are fully justified. In the above context we should mention the efforts to bridge between the low-energy TDHF and high-energy Boltzmann approaches, by either supplementing the TDHF equations with an equation for orbital occupation containing a collision integral or by complementing the equation for one-body density matrix, implicit in TDHF approach, with one for two-body matrix. Realistic calculations for those approaches are limited so far to light systems and low energies .
## The pBUU transport model {#sec:pBUU}
Starting the discussion with the higher energy region, central collisions of heavy ions are laboratory means of bringing nuclear matter to supranormal densities over spatial regions that are large as compared to the range of the nuclear force, for times long enough to allow passage of a signal over the region. In a central collision, moving with the N-N center of mass (longitudinally), a portion of the initial kinetic energy is converted into nuclear compression energy. There will also be a lateral motion during the collision. Particle densities higher than in ordinary nuclei can be reached at beam energies of a few hundred MeV/nucleon. As the beam energy per nucleon increases and becomes comparable with the Fermi energy, elastic N-N collisions begin to contribute to the compression within the system. With a reduced role of the Pauli principle, as the Fermi spheres melt down with the occupations turning to small fractions, the N-N collisions dissipate the longitudinal motion of nucleons within the overlap region of the colliding nuclei, slowing down the passage of the nucleons across that region and causing a density pile-up when additional nucleons flow in. At even higher energies, the colliding nuclei become Lorentz contracted to some degree, before the overlap region develops. Simultaneously, the N-N collision cross sections become forward peaked, allowing the longitudinal motion to persist longer during the reaction. In theory, the highest compression that could be reached in the N-N collision-dominated regime is given by a solution of the hydrodynamic Rankine-Hugoniot equation. This equation is applicable provided the system is large enough to allow the matter in the overlap region to reach complete equilibrium, erasing any memory of the original motion along the beam axis. However, according to transport simulations, the Rankine-Hugoniot limit is not reached even for heaviest nuclei, independent of the beam energy. The present numerical simulations of heavy-ion collisions in the higher-energy regime are based on solution of the semiclassical Boltzmann-Uhlenbeck-Uehling equations using the code of Danielewicz *et al.*, termed recently pBUU. The names of Uhlenbeck and Uehling are often associated with the transport equation set for heavy-ion collisions, as the two authors were the first to modify the original classical transport Boltzmann equation to include the final state Pauli suppression and Bose-enhancement factors . The pBUU model is formulated within relativistic Landau theory in which the state of a system of particles is completely specified when the phase space (Wigner) distribution functions \(f_{\rm X}\) of the particles are given. Here, \(f_{\rm X}({\pmb p}, {\pmb r}, t)/(2\pi)^3 \,\) represents the phase-space probability density for finding a particle of type \(X\), at time \(t\) and position \({\pmb r}\), with momentum \({\pmb p}\) and a specific spin direction. Local density of particles \(X\) is determined from \(\rho_X ({\pmb r},t) = \frac{g_X}{(2\pi)^3} \int \text{d}{\pmb p} \, f_X({\pmb p},{\pmb r},t)\), where \(g_X\) is spin degeneracy. The distributions \(f_{\rm X}({\pmb p}, {\pmb r}, t)\) satisfy a set of the Boltzmann equations (we omit here the UU for simplicity) \[\frac{\partial f_X}{\partial t}+\frac{\partial\epsilon_X}{\partial {\pmb p}}\frac{\partial f_X}{\partial {\pmb r}}-\frac{\partial\epsilon_X}{\partial {\pmb r}}\frac{\partial f_X}{\partial {\pmb p}} = I_X \label{eq:boltz}\] where the single-particle energies \(\epsilon_X\) are derivatives of the total energy \(E\) of the system \[\epsilon_X({\pmb p},{\pmb r},t)=\frac{(2\pi)^3}{g_X}\frac{\partial E}{\partial f_X({\pmb p},{\pmb r},t)}.\] In Eq. [\[eq:boltz\]](#eq:boltz){reference-type="eqref" reference="eq:boltz"} \({\pmb v}_X = \partial \epsilon_X/\partial {\pmb p}\) is the velocity of a particle X and \({\pmb F}_X =-\partial\epsilon_X/\partial {\pmb r}\) is the force acting on the particle. The terms on the l.h.s. of Eq. [\[eq:boltz\]](#eq:boltz){reference-type="eqref" reference="eq:boltz"} account for the changes in \(f_X\) due to the motion of particles in the average mean field produced by other particles. On the r.h.s., the collision integral \(I_X\) that incorporates cross-sections and N-N collision rates governs the modifications of the distributions \(f_{\rm X}({\pmb p}, {\pmb r}, t)\) due to elastic and inelastic N-N collisions and decays caused by short-range residual interactions beyond the mean-field. The inelastic processes, giving rise to resonances and pion production, play a significant role only at beam energies close to 800 MeV/nucleon and their effect is marginal below 400 MeV/nucleon. We employ in-medium cross sections adjusted so that the collision radius does not exceed the typical distance between nucleons in the medium. These cross sections adequately describe stopping observables from heavy-ion collisions, such as linear momentum transfer , unlike the free cross sections. Details on the collision integrals can be found in and and in references therein. At low excitation energies, the N-N collisions are suppressed and the Boltzmann equations reduces to their Vlasov form . Comparison of the pure Vlasov dynamics to that in TDHF were carried out in the past by Tang *et al.* in one dimension and by Wong in three dimensions. Within the pBUU framework the collision dynamics in a mean-field approximation is determined by the dependence of the total energy \(E\) on the phase-space distributions \(f_X\) . The energy \(E\) consists of the covariant volume term and non-covariant gradient correction \(E_{\rm 1}\), and isospin \(E_{\rm T}\) and Coulomb \(E_{\rm Coul}\) terms, all defined in the local rest frame where the nucleon flux vanishes: \[E = \int \text{d} {\pmb r} \, \tilde{e} + E_1 + E_T + E_\text{Coul} \, \label{eq:funct}\] where \(\tilde{e}\) is the volume energy density, \[\begin{aligned}
E_1 & = & \frac{a_1}{2 \rho_0} \int \text{d} {\pmb r} \, \big(\nabla \rho \big)^2 \, ,\\[1ex] E_T & = & 4 \int \text{d} {\pmb r} \, \frac{S_\text{int}(\rho)}{\rho} \, \rho_T^2 \, ,\\[1ex] E_\text{Coul} & = & \frac{1}{4 \pi \epsilon_0} \int \text{d} {\pmb r} \, \text{d} {\pmb r}' \, \frac{\rho_\text{ch}({\pmb r}) \, \rho_\text{ch}({\pmb r}')}{|{\pmb r}-{\pmb r}'|} \,. \end{aligned}\] The expression for the energy density functional \(\tilde{e}\) in [\[eq:funct\]](#eq:funct){reference-type="eqref" reference="eq:funct"} combines computational ease with a relative flexibility in predicting density dependence of the energy per particle at zero temperature and the momentum dependence of the nucleonic mean fields. It reads \[\tilde{e} = \sum_X g_X \int \frac{\text{d}{\pmb p}}{(2\pi)^3} \, f_X({\pmb p}) \, \bigg( m_X + \int_0^p \text{d}p' \, v_X^*(p',\rho) \bigg) + \int_0^\rho \text{d} \rho' \, U_\rho(\rho') \,, \label{eq:etilde}\] where the summation is over particle species X. The in-medium particle velocity in the local frame (denoted by the star superscript) depends on the local density and kinematic momentum as \[v_X^* = \frac{p}{\sqrt{p^2 + m_X^2 \bigg/ \bigg( 1 + c \, \frac{m_N}{m_X} \frac{\rho}{\rho_0} \frac{1}{\big(1+\lambda p^2 /m_X^2\big)^2} \bigg)^2}} \,.\] where m\(_\text{X}\) is free mass of particle \(X\), and \(c\) and \(\lambda\), representing the strength and characteristic scale of the momentum dependence, are adjustable parameters. In [\[eq:etilde\]](#eq:etilde){reference-type="eqref" reference="eq:etilde"}, \(U_\rho\) is a density-dependent contribution to the mean-field taken in an empirical form. The full potential \(U(\rho,{\pmb p})\) consists of \(U_\rho\) and a \(\delta U_p\) which describes the momentum dependence of the mean field : \[U=U_\rho + \delta U_p = (a \, \rho+b \, \rho^\nu)/[1+(0.4 \, \rho/\rho_0)^{\nu-1}] + \delta U_p \,.\] The constants \(a\), \(b\), \(c\) and \(\lambda\) are adjusted to empirical properties of symmetric nuclear matter and characteristics of the momentum dependent mean-field potential (cf. Table [\[tab:1\]](#tab:1){reference-type="ref" reference="tab:1"}). The functional ([\[eq:etilde\]](#eq:etilde){reference-type="ref" reference="eq:etilde"}) parallels the Skyrme density functional ([\[tdhf\]](#tdhf){reference-type="ref" reference="tdhf"}), but contains more elaborate momentum averages than the kinetic energy density used in the Skyrme models. A simple effective mass substitution in the equations of motion, following from the Skyrme functional, would be inadequate in the equations over the range of beam energies explored in heavy-ion collision measurements. The gradient term \(E_{\rm 1}\) allows for the finite range of nuclear forces and its impact, to the lowest order, on the energy and consequently on nuclear densities in the initial state of the reactions. The coefficient \(a_{\rm 1}\) in the gradient term is adjusted to yield realistic diffusivity in the nuclear ground states. The isospin term \(E_{\rm T}\), dependent on the isospin density \(\rho_\text{T}\), affects both the relative proton and neutron densities in the initial state and the densities reached in the compression stage of a heavy-ion reaction. High value of the symmetry energy at supranormal densities may act to reduce the proton-neutron asymmetry at the high densities and, thus, influence the dynamics . As mentioned in Sec. [1](#sec:intr){reference-type="ref" reference="sec:intr"}, the density dependence of the S\(_\text{int}\) (cf. Eq. [\[eq:3\]](#eq:3){reference-type="ref" reference="eq:3"}) is an input to the simulation. We employ four different forms of this density dependence. The first model is the conservative choice of a linear density dependence (labeled S): \(S_\text{int} (\rho) = s_{ 0} \, \rho/\rho_0\). The second model (labeled SM) mimics the density dependence predicted by the SV-bas Skyrme interaction, used in the TDHF calculation (cf. Sec. [2.2](#sec:tdhf){reference-type="ref" reference="sec:tdhf"}): \(S_\text{int}(\rho) = s_{\rm 1} \, \rho/\rho_{ 0}-s_{2} \, (\rho/\rho_{0})^{\rm s_{3}}\) (cf. Table [\[tab:1\]](#tab:1){reference-type="ref" reference="tab:1"}). Although these models produce fairly similar values of the symmetry energy at normal density \(\sim\[31 \, \text{MeV}\), the slope parameters, \(L = 3 \rho_0 \, (\partial S/\partial \rho) |_{\rho_0}\), are significantly different: \(L = 85 \, \text{MeV}\) for S and \(L= 31 \, \text{MeV}\) for the SM model. The hybrid third (SMS) and fourth (SSM) models are constructed using S and SM parameterizations, in order to test competing impacts of the low-density and high-density behavior of \(S_\text{int}(\rho)\). The third SMS parametrization behaves like SM at low densities and like S at high. In the SSM parametrization, the order is reversed, i.e. S is followed at low and SSM at high densities. In each case, the a smooth transition between the S and SM behaviors takes place at a mildly subnormal density \(\rho \lesssim \rho_0\), where the S and SM parametrizations cross, cf. Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} where we illustrate the density dependence of the total symmetry energy as predicted by the four models. The initial state of nuclei entering a reaction is determined by solving the Thomas-Fermi (TF) equations yielded by the requirement that the total energy ([\[eq:funct\]](#eq:funct){reference-type="ref" reference="eq:funct"}) is minimal in the ground state. The set of Boltzmann equations ([\[eq:boltz\]](#eq:boltz){reference-type="ref" reference="eq:boltz"}) for nucleons, baryon resonances and pions is solved using the test-particle Monte-Carlo method. Given that this work focuses on finding maximum particle densities over position and time, any statistical fluctuations in the Monte-Carlo calculations outside of the model framework may be of concern. To minimize possible effects of these fluctuations on deduced maxima, we use a rather large number of test-particles per physical particle, \({\mathcal N}_Q = 3000\). The pBUU model depends on a number of adjustable parameters as introduced above and summarized in Table [\[tab:1\]](#tab:1){reference-type="ref" reference="tab:1"}. The particular combination of parameter values yields fairly conservative properties of nuclear matter, but that parameter set can be varied to explore sensitivity of the progress of a reaction and of reaction observables to the characteristics of the matter.
## The TDHF model {#sec:tdhf}
The method for treatment of low-energy heavy-ion collisions that currently best combines realism with flexibility and convenience is the Time-Dependent-Hartree-Fock (TDHF) theory, a nonperturbative approach that allows description of multi-nucleon transfer. At low beam energies, the reaction mechanism of heavy-ion collisions is rather different from that discussed in Sec. [2.1](#sec:pBUU){reference-type="ref" reference="sec:pBUU"}. At energy sufficient to overcome the Coulomb barrier, the colliding system goes through two main phases, fusion and full overlap of the target and projectile, followed by disintegration into two or more fragments. The typical nucleon mean-free-path at low energies exceeds the size of the composite system. This is in contrast with strongly excited systems produced at higher beam energies, where nucleon mean-free-path shrinks below the system size and continues to decrease as the beam energy increases. Unlike in high energy collisions when any equilibration has a hard time completing before the reaction ends, in low energy collisions several types of equilibrations may occur, each on a different time scale. At early stages, a fast charge (chemical) equilibration takes place, which arises from a motion of nucleons near the Fermi surface. The basic mechanism of this equilibration is understood in terms of extension of the single particle motion from one nucleus to another, following the lowering of the potential barrier between the two colliding nuclei after contact. This process can be very fast, having a typical time scale of \(\sim\]10^{\rm-22} \, \text{s} \equiv 30 \, \text{fm}/c\) . The charge equilibrium is quite important because it prevents production of exotic fragments with extreme proton-neutron asymmetry. Relatively longer time \(\lesssim\[10^{\rm -21} \, \text{s} \equiv 300 \, \text{fm}/c\) is needed for formation of a composite system in which the single-particle motion becomes synchronized under the influence of the nuclear force to yield collective oscillations. Density equilibration, in which the matter forms a bound state with interior density around the normal value \(\rho_{\rm 0}\), may also occur, but is not well established its time scale is difficult to estimate. It is usually studied together with charge and momentum equilibration. Momentum equilibrium that represents a balance between nuclear and Coulomb forces, is expected to take about \(10^{\rm-20} \, \text{s}\). Slower equilibration processes, such as thermal equilibration, and some specific fission and/or decay processes of the composite system, can also take place over times of the order of \(10^{\rm-15} \, \text{s}\) in low-energy collisions, but these processes are beyond the scope of this work. The TDHF dynamics presented in this work has been obtained from Sky3D code designed to solve both static and time-dependent Hartree-Fock equations within a general three-dimensional geometry. Detailed description of the physics and handling of the code can be found in Ref. . The model is based on a Skyrme energy functional of the general form \]E_{\rm tot} =E_\text{kin}+E_\text{Skyrme}+E_{\rm Coul}+E_{\rm pair}+E_{\rm corr} \,, \label{tdhf}\(\) where \(E_\text{kin}\) is the total kinetic energy and \(E_\text{Skyrme}\) stands for terms related to the Skyrme force. The other terms are additions, not treated self-consistently. They represent the Coulomb and pairing forces and effects of correlations in the mean-field, such as a correction for center of mass motion, a rotational correction for deformed nuclei and a correction for correlations arising from all low-energy quadrupole degrees of freedom in soft nuclei. In the present work we employ the SV-bas interaction and neglect the pairing and correlation terms in ([\[tdhf\]](#tdhf){reference-type="ref" reference="tdhf"}). SV-bas is a set of 15 parameters, fitted to 70 independent pieces of data on ground-state binding energies of magic and semi-magic even-even nuclei and to over 75 pieces of experimental data on rms charge and diffraction radii, a neutron skin value and single-particle energies. In addition, constraints on properties of nuclear matter, namely the value of incompressibility set at \(K = 234\, \text{MeV}\), effective mass \(m^*/m = 0.9\), the symmetry energy coefficient \(a_\text{sym} = 30 \, \text{MeV}\), and the enhancement factor in the Thomas-Reiche-Kuhn sum rule at \(\kappa = 0.4\) were imposed in the fit. For some specific examples of physics application for the SV-bas interaction cf. Ref. . Both TDHF and pBUU models are mean-field approaches to particle dynamics in nuclear collisions. They differ in that TDHF is a quantum-mechanical framework utilizing single-particle wavefunctions rather than phase-space probability densities in the semi-classical pBUU. In TDHF, the initial state of a collision is obtained from static Hartree-Fock (HF) equations as compared to the TF approach in pBUU. A set of time-dependent Schrödinger equations is solved self-consistently to obtain wavefunctions describing the time evolution of a colliding system, and yielding spatial proton and neutron densities at any time and location. The TDHF method is appropriate for simulating low-energy heavy-ion reactions at incident energy up to about \(40\, \text{MeV/nucleon}\). The main reasons for such a limitation is that TDHF does not include any correlations beyond the mean field or N-N collisions which would affect the dynamics. The advantage of the microscopic TDHF approach is that it includes shell effects, most important in low energy reactions, which are not included in the semiclassical pBUU and typically in the molecular dynamics models where, however, N-N collisions are easily incorporated.
# Results and Discussion {#sec:res}
In this section we illustrate the main results of the calculations in the whole beam energy range. We point out similarities and differences in the pBUU and TDHF dynamics across the energy region where both approaches are expected to work. The proton-neutron asymmetries reached at high densities are examined in order to distinguish contributions of asymmetries inherited from the initial state and coming from high-density behavior of the symmetry energy.
## Static Densities {#sec:static}
Figure [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"} shows the density distributions of protons and neutrons, as a function of distance from the center of a nucleus, obtained by solving static TF (in pBUU) and HF (in TDHF) equations. In the absence of shell effects within the TF theory, the resulting densities have a smoother dependence on distance and charge and mass number as compared to the HF theory. Since tunneling effects are not incorporated to the TF theory, the TF densities lack the tails falling off exponentially with distance that are evident in the HF densities. Following general expectations, the central neutron densities are similar to proton densities in case of the N = Z nuclei, differing by less than 10% for both \(^{40}\)Ca and \(^{100}\)Sn in the HF and TF models, cf. Fig. 2. The HF model predicts the magnitude of both densities being somewhat higher as compared with the TF model in \(^{40}\)Ca. In the N \(\neq\) Z systems, the HF predictions of magnitude of the central densities are also higher than in the TF model. The central proton-neutron asymmetry is also model dependent. It is about a factor of two bigger for \(^{120}\)Sn and somewhat smaller for \(^{48}\)Ca in HF as compared to the TF calculation, These results can be related to differences in physics included in the HF and TF models. The presence of shell effects in the HF model and its microscopic nature are expected to make the approach more suitable for calculation of static nuclei. Thus, for example, the lack of shell effects and generally the approximate treatment of densities in the TF model can be the reason that, while in \(^{120}\)Sn the difference between the central proton and neutron densities is higher in HF than TF, in \(^{48}\)Ca the differences are reversed, see Fig. 2. The shell effects influence also another significant factor that plays a role, the density distributions at the surface as compared to the center of a nucleus. In the lighter Ca systems the ratio of surface to volume is larger that in the Sn systems. As the symmetry energy is lower at lower densities, the surface areas are energetically favorable for development of relative asymmetry as compared to the center where the density is higher. Thus the proton-neutron imbalance at the center is expected to be lower in Ca than in Sn nuclei. As discussed more in detail in Sec. [3.3](#sec:pnasym){reference-type="ref" reference="sec:pnasym"}, the static proton-neutron asymmetry and its development at subnormal densities affect significantly the asymmetry at supranormal densities and, thus, has to be determined as well as possible in order to make trustworthy reaction predictions. The choice of a symmetry energy model from our pool affects very little the central proton and neutron densities in the TF calculations. We illustrate this effect showing results for S and SM in Fig. [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}. The results for SSM and SMS models are not included as they are practically identical to those obtained with S and SM models, respectively. This is to be expected as there is very little difference within the pairs of the models at sub-saturation densities (cf. Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"})
## Time evolution of a collision: contrasting approaches, system sizes and energies {#sec:time}
Detailed study of time evolution of proton and neutron densities as a function of beam energy is essential to our particular goal of extracting maximal values of the densities during a central collision. Two representative cases of time evolution are illustrated in Fig. [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"} for asymmetric \(^{\rm 40}\)Ca + \(^{\rm 48}\)Ca and \(^{\rm 100}\)Sn + \(^{\rm 120}\)Sn projectile+target systems. Results yielded by the pBUU dynamics with the S model for the symmetry energy, for beam energy \(16 \leq E_{\rm beam} \leq 400 \, \text{MeV/nucl}\) are complemented by the outcome of the TDHF calculation for \(16 \leq E_{\rm beam} \leq 40 \, \text{MeV/nucl}\), the beam energy region where the TDHF model is applicable. The time-dependence of maximal densities in Fig. [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"} exhibits some common features characteristic for all the systems we investigate, irrespectively of the approach. First, the neutron and proton density tends to maximize at about the same time, i.e. the maximum compression of neutrons and proton matter is synchronized. Second, the time elapsed since early contact between the colliding nuclei till the absolute maximum is reached is beam energy dependent, dropping as the energy is increased. In more energetic collisions, modeled only in pBUU, the maximal density drops in a rather dramatic fashion after the maximum compression has been reached. This drop corresponds to a system vaporization. The times for maximal compression and vaporization change little for the Ca systems between \(E_\text{beam} = 200\) and \(400 \, \text{MeV/nucl}\). In the Sn systems more noticeable delays develop with energy. Results at \(E_{\rm beam} \leq 40 \, \text{MeV/nucl}\) offer not only exploration of the lower beam energy region but also a comparison of the pBUU and TDHF model predictions. As illustrated in Fig. [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}, in both the Ca and Sn systems the first maximum, typically absolute, is followed by a series of oscillations with amplitudes decreasing faster with time in pBUU than in TDHF. The nature of these oscillations in TDHF is illustrated in Fig. [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}, which displays contour plots of proton and neutron densities at sample times in the \(^{100}\)Sn+\(^{120}\)Sn collisions at the energies of 16 and \(40 \, \text{MeV/nucl}\). At lowest energies, the oscillations represent motion of a fused system. At somewhat higher energies, the systems tend to fission after reaching the maximum compression and the oscillations represent motion of separated fission fragments. The out-of-phase oscillations in neutron and proton densities, particularly visible in TDHF, are related to waves propagating across the colliding system that are fast compared to the overall system evolution. In a projectile-target asymmetric system, the oscillation starts with a charge equilibrating wave propagating across the fusing nuclei. This wave is combined with a spontaneous dipole oscillation along the direction of the reaction that adds, with an *ad hoc* phase, to the existing proton and neutron density differences. The shape relaxation, combined with Coulomb repulsion, enhances multipolarity of the isovector waves, as demonstrated in Fig. [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}. Here we see fission, following the early fusion, traps instantaneous asymmetries at both ends of the system within the separating fragments which subsequently re-contract. This relatively complicated and weakly damped isovector dynamics in TDHF leads to effects such as prediction of a stronger compression of protons than neutrons in the second peak of the oscillation (cf. Fig. [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}). Comparing further the Ca and Sn systems in the pBUU and TDHF models, relaxation after the first compression takes longer in pBUU than in TDHF. These times are almost the same for Ca and Sn systems in TDHF, but in pBUU the Ca system relaxes faster than the Sn system. Although the overall pattern of the density evolution is not very different between the approaches at \(E_{\rm beam} = 16 \, \text{MeV/nucl}\), for both Ca and Sn, there is a markable difference in predictions of the TDHF and pBUU models for \(E_{\rm beam} = 40 \, \text{MeV/nucl}\), both in amplitudes of the first compression and the onset of oscillations. This effect may be related to the collisional equilibration taking place in pBUU but not TDHF and to overall stronger damping of collective waves in pBUU than TDHF.
## Maximum densities and the proton-neutron asymmetry {#sec:pnasym}
Using the maximal densities \(\rho_p^\text{max}\) and \(\rho_n^\text{max}\), extracted from data on time evolution of a collision as detailed in Sec. [3.2](#sec:time){reference-type="ref" reference="sec:time"}, we construct the asymmetry \(\delta = (\rho_n^\text{max}-\rho_p^\text{max})/(\rho_n^\text{max} + \rho_p^\text{max})\) as a function of system and beam energy. Although, strictly speaking, the maximal densities may not be reached at exactly the same time, or at the same location, we find that \(\delta\) gives a good overall representation of the asymmetry in the densest matter in a collision, even in systems with different projectile and target initial asymmetries. The proton and neutron densities fluctuate as a function of beam energy in pBUU as a consequence of the Monte-Carlo nature of the calculations and due to the propagating waves with a strong isovector component in TDHF. The fluctuations can obscure gradual changes, particularly in a differential quantity such as the asymmetry \(\delta\). To circumvent the issue of fluctuations, in addition to extracting individual values of the maximum densities, we approximate them at each beam energy \(x\) using a smoothing function \(y = a_{\rm 0}+a_{\rm_1}\log(1+a_{\rm 2} x)\). These interpolated densities, shown in Figs. [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}--[\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"} for pBUU and in Fig. [\[fig:8\]](#fig:8){reference-type="ref" reference="fig:8"} for TDHF, are then used to calculate \(\delta\). The parameters \(a_{\rm 0}\), \(a_{\rm_1}\) and \(a_{\rm 2}\), summarized in the Supplemental Material, show reasonable consistency in all related cases. It follows that the interpolation function is an adequate representation of the raw data. We note that the error in the fitting is within a few percent in all cases and does not affect conclusions of the analysis. Figures [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}--[\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"} illustrate the evolution of the interpolated maximum proton and neutron densities with increasing beam energy for the Ca and Sn systems, respectively, as simulated in pBUU, using four models of the density dependence of the symmetry energy S, SM, SMS and SSM. The results indicate that \(\rho_{\rm n}^{\rm max}\) does not exceed \(\sim\)`<!-- -->`{=html}1.3\(\times\rho_{\rm 0}\) for Sn systems, with the \(\rho_{\rm p}^{\rm max}\) being below \(\sim\)`<!-- -->`{=html}1.1\(\times\rho_{\rm 0}\). The results for the lighter Ca systems are marginally lower. In the Vlasov approximation (cf. Fig. [\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}), in which the N-N collisions are eliminated, the maximum densities are lower. The highest maximum particle number density, as resulting from all models used in this work for Ca and Sn systems and beam energies, is \(\sim\)`<!-- -->`{=html}2.5\(\times\rho_{\rm 0}\) at 800 MeV/nucl. The total nucleon densities normalized to the initial static values (compression ratios) are given in Tables [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"} and [\[tab:3\]](#tab:3){reference-type="ref" reference="tab:3"} for pBUU and TDHF, respectively. The ratios generally rise with increasing beam energy. Interestingly, the compression ratios depend on the size of a system. Within the elements, the pBUU model predicts the highest compression ratios for the heaviest (symmetric) Ca and Sn systems, respectively, cf. Table [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"}. The TDHF model yields the same result for Ca, but in the Sn systems the situation is reversed and \(^{\rm 100}\)Sn + \(^{\rm 100}\)Sn combination yields the highest compression. This is likely related to a strong shell effect, enhancing initial density in \(^{120}\)Sn, that apparently dissolves during the collision, but is used in the normalization. In the region of overlap of the applicability of pBUU and TDHF, we consistently find higher compression ratios in the pBUU model. The difference is even more obvious from comparison of TDHF to pBUU in the Vlasov mode, conceptually closer to TDHF, as neither of these models includes elementary N-N collisions. It seems that the N-N collisions are detrimental to compression at lower energies, but they enhance it at higher energies, cf. Table [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"}. The TDHF model yields systematically lower compression in the beam energy region below \(40 \, \text{MeV/nucl}\), than the pBUU model in the same region, reaching maximum compression ratio \(\sim\)`<!-- -->`{=html}1.4 at 40 MeV/nucl for Sn and \(\sim\)`<!-- -->`{=html}1.3 for the Ca systems (cf. Table [\[tab:3\]](#tab:3){reference-type="ref" reference="tab:3"} and Fig [\[fig:8\]](#fig:8){reference-type="ref" reference="fig:8"}). Higher compression ratios are predicted above the beam energy \(50 \, \text{MeV/nucl}\), reaching \(\sim 2.5\) at \(800 \, \text{MeV/nucl}\) in the pBUU model with N-N collisions included. A remarkable feature of the asymmetries \(\delta\), shown in the bottom panels of Figs. [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}--[\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}, is that they evolve with the beam energy more rapidly at low energies than at high energies. This effect may be understood in terms of the pace of the dynamics. At lower beam energies, the dynamics is more adiabatic, with density profiles and asymmetry values undergoing broader adjustments before maximal densities being reached. At high energies, the central region of the system becomes compressed in a shock manner and the asymmetry represents closer the matter that suddenly has fallen into that region. In the Ca systems, a significant portion of proton-neutron imbalance is pushed out to the surface, so asymmetries at maximum compression are much reduced, almost by a factor of two, compared to the total asymmetry expected from the total number of protons and neutron in the system. The asymmetries at maximum densities in the Sn systems are higher, both because of the lower surface area to volume ratio and as a consequence of stronger Coulomb repulsion, disfavoring protons in the central region of the system. Comparing results with and without N-N collisions (in the Vlasov dynamics), in Figs. [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}--[\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}, we see that when the N-N collisions are active, they help to trap the asymmetry in the high-density regions. The N-N collisions have also a different impact on asymmetry in dependence on the size of a system. The mean-free-path between N-N collisions is shorter, relative to system size, for Sn than for Ca, making the collision more effective in trapping the isospin excess in high density regions, thus increasing the asymmetry coming from the initial proton-neutron distribution and Coulomb repulsion. In comparing asymmetries \(\delta\) reached in TDHF and pBUU, in the overlap region for the approaches, Fig. [\[fig:8\]](#fig:8){reference-type="ref" reference="fig:8"} and Figs. [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}--[\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}, it is observed that the asymmetry for maximal densities is generally higher in TDHF than in pBUU, particularly at low energy. This effect seems to be related to the undamped waves in TDHF that, combined with fission, increase asymmetry in one fragment at the cost of the other. The fragment separation enhances the maximal neutron density less than the maximal proton density, due to the increased Coulomb repulsion in the fragment with reduced asymmetry, i.e. extra charge number per neutron. With increasing incident energy the predictions of TDHF and of pBUU for the SM energy parametrization get closer to each other. The apparent impact of the sudden dive of matter into a high density region, on the high-density proton-neutron asymmetry in collisions at high energy, raises questions regarding the feasibility of probing the high-density symmetry energy in the laboratory. Comparing the high-energy asymmetries in the panels labeled S and SM in Figs. [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"} and [\[fig:6\]](#fig:6){reference-type="ref" reference="fig:6"}, it could be argued that the simulation with higher symmetry energy at high density, i.e. S model cf. Fig. [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}, yields lower asymmetries as the excess neutrons are expelled from the high-density region . However, it could also be argued that the difference in the high-density asymmetries observed at high beam energies, resulting from S and SM calculations, stems exclusively from the difference in interior asymmetries within the nuclei at the start of a collision, dooming the possibility of testing the high-density symmetry energy in heavy-ion collisions. In order to examine a possible impact the high-density symmetry energy may have, we constructed the hybrid symmetry-energy models, SSM and SMS, for which the initial states are practically identical with the S and SM models, respectively, but the high-density behavior of the symmetry energy is swapped in-between the models. Comparing S with SSM and SM with SMS for Ca and, even more importantly, for Sn, we see that the high-density behavior of the symmetry energy does make a difference. If the symmetry energy is switched at high density from stiff (raising fast with increasing density) to soft raising slowly with increasing density), i.e. S to SSM, the high-energy asymmetry increases. On the other hand, if the models are reversed, i.e. SM to SMS, the high-energy asymmetry decreases. It seems that the SMS model is more effective in loosing asymmetry than SSM model, particularly in the lighter Ca systems where the reaction lasts for a shorter time. In other words, it is easier to lose asymmetry than to reconstitute it after it has been depleted in the dynamics when the high-density symmetry energy is stiff and *vice versa*. Overall, the experimentation with the symmetry energy models shows that both, the initial distribution of asymmetry in the nuclei and the high-density behavior of the symmetry energy impact the high-density asymmetry. To reveal the effect of the initial distribution of asymmetry on the high-density behavior of symmetry energy, the static nucleon distributions in the target and projectile nuclei must be modeled consistently in the simulations. For the most proton-neutron asymmetric systems, \(^{48}\)Ca + \(^{48}\)Ca and \(^{120}\)Sn + \(^{120}\)Sn, we find changes in the high-density asymmetry of up to \(\sim 20\)% in the lighter system and up to \(\sim 15\)% in the heavier when the symmetry energy is varied, cf. Figs. [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"} and [\[fig:6\]](#fig:6){reference-type="ref" reference="fig:6"}. Paradoxically, the higher sensitivity to symmetry energy in the lighter system seems to be related to the combined ratios of surface thickness in the initial state and the mean-free-path to system size in the dynamics. Importantly if the nuclear systems were very large, the asymmetry at high density would neither change due to low-nor high-density behavior of the symmetry energy. This is because the neutron-proton imbalance would have practically nowhere to move from the interior of matter neither in the initial nor in the compressed state. With this, while the general expectation is that heavier systems are better suited for testing EOS, in the particular case of supranormal symmetry energy, lighter systems may be more suitable, when the dedicated observables are tied to the value of asymmetry reached at high densities.
# Conclusions and outlook {#sec:concl}
We presented results of the first systematic examination of maximal values and other details in neutron and proton distributions, within the initial state and throughout the course of head-on heavy-ion collisions across the beam energy range of up to \(800 \, \text{MeV/nucl}\), in two common simulation approaches, TDHF and pBUU. The projectile-target symmetric systems \(^\text{40}\)Ca +\(^\text{40}\)Ca, \(^\text{48}\)Ca +\(^\text{48}\)Ca, \(^\text{100}\)Sn +\(^\text{100}\)Sn and \(^\text{120}\)Sn + \(^\text{120}\)Sn were selected for this study, as well as the projectile-target asymmetric systems \(^\text{40}\)Ca +\(^\text{48}\)Ca and \(^\text{100}\)Sn +\(^\text{120}\)Sn. Within the pBUU simulations, two empirical forms of symmetry energy, characterized by \(L=85 \, \text{MeV}\), termed S, and \(L=31 \, \text{MeV}\), termed SM, and two hybrid forms, termed SSM and SMS were employed. Simulations with suppressed collisions, i.e. in the Vlasov mode were also performed for completeness. TDHF calculations were carried out in the beam energy range of up to \(40 \, \text{MeV/nucl}\), with the SV-bas Skyrme parametrization that yields the symmetry energy close to that in the SM parametrization within pBUU. Maximum proton and neutron densities \(\rho_{\rm p}^{\rm max}\) and \(\rho_{\rm n}^{\rm max}\), reached in the course of the collisions, were identified from analysis of the time evolution of \(\rho_{\rm p}\) and \(\rho_{\rm n}\) during the course of the collision and used to yield the proton-neutron asymmetry \(\delta = (\rho_{\rm n}^{\rm max}-\rho_{\rm p}^{\rm max})/ (\rho_{\rm n}^{\rm max} + \rho_{\rm p}^{\rm max})\) as a function of beam energy and all target/projectile combinations considered in this work. In TDHF, the initial densities, as calculated the HF model, are characterized by significant shell oscillations unlike the TF starting densities for pBUU. The oscillations, combined at low energies with weakly damped collective waves with significant isovector components, enhance density variations across the system and time. When a colliding system fissions, the variations may become trapped in opposite fragments, leading to a larger proton-neutron imbalance than in the system before fission. In combination with Coulomb effects, this effect enhances the proton-neutron asymmetry at the maximal recorded densities. As the beam energy increases, the impact of the initial-state shell effects decreases and the TDHF results become similar to pBUU. The pBUU simulations indicate that the maximum total densities reached in reactions studied in this work are \(\sim 2.5 \, \rho_0\) (\(0.4 \, \text{fm}^{\text-3}\)) are not significantly dependent on the initial state, beam energy, system size and a model of the symmetry energy. However, the proton-neutron asymmetry at maximum densities does depend on these parameters. The simulation in this work indicates without any doubt that the initial asymmetry of all systems studied plays an important role in a reaction. The asymmetry decreases after the impact, changes slowly in the course of a collision as a function of beam energy, but never exceeds the original value. For example, for the most neutron-proton asymmetric systems, i.e. \(^{48}\)Ca + \(^{48}\)Ca and \(^{120}\)Sn + \(^{120}\)Sn, both with the overall initial asymmetry of 0.17, we find high-density asymmetries at high beam energies in the range of (0.09--0.12) for the lighter system and (0.12--0.15) for the heavier, depending on the symmetry energy model employed. It has become clear that in predicting the impact of the density dependence of symmetry energy on high-density asymmetry, the initial state must be modeled consistently with the dynamics that leads to development of the compressed matter. If the rate of change of the initial asymmetry could be measured as a function of the beam energy, it would give information of the high-density behavior of the symmetry energy. When changing the symmetry energy, we find somewhat more impact on the high density asymmetry in the lighter asymmetric systems than heavier. To summarize, the main outcome of our work is the identification of importance of the sub-normal density dependence of the symmetry energy in determination of its form in supra-normal density. There is an effort in the community to learn about the supra-normal symmetry energy, with different variants explored in theoretical simulations. One practical issue is that changes in the symmetry energy model in the supra-normal region, smoothly matched to a model in the sub-normal region, inevitably affects the initial state of the collision. Our simulation reveals that only about half of the proton-neutron asymmetry at super-normal densities comes directly from behavior of the supra-normal symmetry energy and the remaining half has its origin in changes in the initial state. This result puts to forefront the need of a proper initialization of the nuclei, but also brings up the question of the shell effects that impact initial proton and neutron densities, but cannot be consistently incorporated into semiclassical transport. In exploring the density dependence of the symmetry energy in central reactions, lighter systems may be as useful as heavier. | {'timestamp': '2017-06-07T02:02:52', 'yymm': '1706', 'arxiv_id': '1706.01582', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01582'} |
null | null |
null | null |
# Introduction
The usual way of studying the geometry of curves is by means of the well known Frenet frame. Such a frame is rich in geometric informations, but since its principal normal always points to the center of curvature, it may result in unnecessary rotation and then making its use unsuitable in some contexts. In this respect, the consideration of *rotation minimizing frames* \(\{\mathbf{t},\mathbf{n}_1,\mathbf{n}_2\}\) (RM frames, for short) is of special interest: the basic idea is that \(\mathbf{n}_i\) rotates only the necessary amount to remain normal to the tangent \(\mathbf{t}\). Due to their minimal twist, RM frames are of fundamental importance in many branches, such as in camera and rigid body motions, visualization and deformation of tubes, sweep surface modeling, and in differential geometry as well, just to name a few. While the computation of a Frenet frame can be easily done, this is not the case for RM frames. Since the inability to compute RM frames would impose severe restrictions in their applications, numerical methods and special curves for which such frames can be computed were investigated. In the numerical arena, we can mention the rotation, projection, and double reflection methods (for more details on numerical approaches, see and references therein). On the other hand, an important class of curves is the so called Pythagorean-hodograph (PH) curves and its subset of polynomial curves that admit rational RM frames: the possibility of using rational function integration in order to exactly compute RM frames is of major importance and leads to many computationally attractive features. Spherical curves constitute an important class for which it is possible to exactly find RM frames. The goal of this work is to study some properties of spherical curves using these frames. We show how to find the angle between the principal normal and an RM vector (the derivative of this angle gives the torsion) for an at least three times differentiable space curve. This is done by using a convenient expression for the curvature function and torsion of a spherical curve. Subsequently, we show how to generalize these expressions for a generic curve by using the concept of osculating spheres, i.e., we describe the behavior of the curvature and torsion of a curve near an osculating sphere using spherical analogs. We also furnish a characterization of spherical curves by using osculating spheres described in terms of RM frames. In addition, we also address in this work the problem of characterizing those curves whose position vector lies, up to a translation, on a (moving) plane spanned by the unit tangent and a rotation minimizing vector and prove that they are precisely the plane curves. This problem has to do with the more general question of studying curves that lie on a given (moving) plane generated by two chosen vectors of a moving trihedron, e.g., one would define osculating, normal or rectifying curves as those curves whose position vector, up to a translation, lies on their osculating, normal or rectifying planes, respectively. It is known that (i) osculating curves are precisely the plane curves[^1], (ii) normal curves are precisely the spherical curves[^2], and (iii) rectifying curves are precisely geodesics on a cone ([^3]): these investigations can be generalized to higher dimensions, to moving frames adapted to surfaces or to other ambient spaces as well. The remaining of this work is divided as follows. In section 2 we introduce RM frames and some geometric background. In section 3 we characterize spherical curves and compute their RM frames. In section 4 we describe the local geometry of curves in terms of spherical analogs by using osculating spheres. In section 5 we establish a characterization of planar curves using RM frames and, finally, in section 6 we present our concluding remarks.
# Preliminaries
Let us denote by \(\mathbb{E}^3\) the 3D Euclidean space, i.e., \(\mathbb{R}^3\) equipped with the standard metric \(\langle x,y\rangle=\sum_{i=1}^{\,3}x_iy_i\). Given a regular curve \(\alpha:I\rightarrow \mathbb{E}^3\), i.e., \(\langle\alpha',\alpha'\rangle\not=0\), the usual way to introduce a moving frame along it is by means of the Frenet frame \(\{\mathbf{t},\mathbf{n},\mathbf{b}\}\): \[\mathbf{t} = \frac{\alpha'}{\Vert\alpha'\Vert},\,\mathbf{b}=\frac{\alpha'\times\alpha''}{\Vert\alpha'\times\alpha''\Vert},\mbox{ and }\,\mathbf{n}=\mathbf{b}\times\mathbf{t},\] whose equations of motion are \(\mathbf{t}'=\Vert\alpha'\Vert\kappa\,\mathbf{n}\), \(\mathbf{n}'=\Vert\alpha'\Vert(-\kappa\,\mathbf{t}+\tau\,\mathbf{b})\), and \(\mathbf{b}'=-\Vert\alpha'\Vert\tau\,\mathbf{b}\): \(\kappa=\Vert\alpha'\times\alpha''\Vert\Vert\alpha'\Vert^{-3}\) and \(\tau=\langle\alpha',\alpha''\times\alpha'''\rangle\Vert\alpha'\times\alpha''\Vert^{-2}\) are the curvature and torsion, respectively. For simplicity, we assume that all the curves are parametrized by an arc-length \(s\), i.e., \(\Vert\alpha'(s)\Vert=1\), unless otherwise stated. In addition, we say that \(\alpha\) is a *twisted curve* if \(\kappa>0\) and \(\tau\not=0\). A point \(\alpha(s^*)\) with \(\kappa(s^*)=0\) is called an *inflection point*. Despite the importance of Frenet frames in geometric considerations, in some applications their use is unsuitable since the normal vectors \(\mathbf{n}\) and \(\mathbf{b}\) rotate around each other. So, one should consider other adapted orthonormal moving frames \(\{\mathbf{t}(s),\mathbf{n}_1(s),\mathbf{n}_2(s)\}\) along \(\alpha(s)\) with the additional property of \(\mathbf{n}_i\) rotating around the unit tangent \(\mathbf{t}\) only. The equation of motion of such a *rotation minimizing* (RM) moving frame is \[\frac{{\rm d}}{{\rm d}s}\left( \begin{array}{c} \mathbf{t}\\ \mathbf{n}_1\\ \mathbf{n}_2\\ \end{array} \right)=\left( \begin{array}{ccc} 0 & \kappa_{1} & \kappa_{2}\\-\kappa_{1} & 0 & 0\\-\kappa_{2} & 0 & 0\\ \end{array} \right)\left( \begin{array}{c} \mathbf{t}\\ \mathbf{n}_1\\ \mathbf{n}_2\\ \end{array} \right).\label{eq::BishopEqs}\]
By writing \(\mathbf{n}_1=\cos\theta\,\mathbf{n}-\sin\theta\,\mathbf{b}\) and \(\mathbf{n}_2=\sin\theta\,\mathbf{n}+\cos\theta\,\mathbf{b}\), the coefficients \(\kappa_1,\,\kappa_2\) relate with the curvature function \(\kappa\) and torsion \(\tau\) according to \[\left\{ \begin{array}{c} \kappa_1(s) = \kappa(s)\,\cos\theta(s)\\[4pt] \kappa_2(s) = \kappa(s)\,\sin\theta(s)\\[4pt] \theta'(s)= \tau(s)\\ \end{array} \right.\.\label{eq::RelBetweenRMandFrenet}\] An advantage of an RM moving frame is that it can be globally defined even if \(\kappa=0\) at some points. In addition, RM frames are not uniquely defined, since any rotation of \(\mathbf{n}_i\) on the normal plane still gives an RM field, i.e., the angle \(\theta\) is well defined up to an additive constant. Nonetheless, the prescription of curvatures \(\kappa_1,\kappa_2\) still determine a curve up to rigid motions. Finally, we say that two regular curves curves \(\alpha\) and \(\beta\) in \(\mathbb{E}^3\) have a *contact of order \(k\)* if \(\alpha(s_0)=\beta(s_0^*)\) and all the higher order derivatives, up to order \(k\), also coincide: \(\alpha^{(i)}(s_0)=\beta^{(i)}(s_0^*)\) for \(1\leq i \leq k\). For example, the tangent line has a contact of order 1 with its reference curve, while the osculating circle has a contact of order 2: at an inflection point the tangent line has a contact of order 2 with its reference curve, so we may say that the osculating circle at this point has an infinity radius of curvature. Further, we say that a curve \(\alpha\) and a surface \(\Sigma\) has a contact of order \(k\) if there exists a curve in \(\Sigma\) which has a contact of order \(k\) with \(\alpha\) and all the other curves has a lower, or equal, order of contact[^4]. For example, the osculating plane, i.e., the plane spanned by \(\{\mathbf{t},\mathbf{n}\}\), has a contact of order 2 with its reference curve, while the osculating sphere has a contact of order 3: at a zero torsion point the osculating plane has a contact of order 3, so we may say that the osculating sphere at this point has an infinity radius. At a twisted point, the center and radius of the osculating sphere are respectively given by \[P_S(s_0) = \alpha+\frac{1}{\kappa}\,\mathbf{n}+\frac{1}{\tau}\frac{{\rm d}}{{\rm d}s}\left(\frac{1}{\kappa}\right)\mathbf{b}\,\mbox{ and }\,R_S(s_0) = \sqrt{\frac{1}{\kappa^2}+\frac{1}{\tau^2}\left[\frac{{\rm d}}{{\rm d}s}\left(\frac{1}{\kappa}\right)\right]^2}\,\.\]
# The geometry of spherical curves
The study of spherical curves leads to many interesting geometric investigations and in this respect the study of the osculating spheres may furnish valuable informations about the geometry of a generic regular curve. The next two sections will be devoted to the investigation of the geometry of spherical curves and osculating spheres in the study of more general curves.
## Characterizing spherical curves
Interestingly, RM frames allows for a simple characterization of spherical curves. Indeed[^5]
Here we furnish a proof of the above result by using osculating spheres. First, let us describe its parametrization by using an RM frame. Indeed, let us write \[P_S(s_0)=\alpha(s_0)+\beta_0\mathbf{t}(s_0)+\beta_1 \mathbf{n}_1(s_0)+\beta_2\mathbf{n}_2(s_0).\] Now, defining a function \(g(s)=\langle P_S-\alpha(s),P_S-\alpha(s)\rangle-r^2\), we have \[\begin{aligned}
g ' & = &-2\langle P_S-\alpha,\mathbf{t}\rangle =-2\beta_0\,,\\ g'' & = & 2\langle\mathbf{t},\mathbf{t}\rangle-2\langle P_S-\alpha,\kappa_1\mathbf{n}_1+\kappa_2\mathbf{n}_2\rangle =-2(-1+\kappa_1\beta_1+\kappa_2\beta_2)\,,\\ g''' & = &-2\langle P_S-\alpha,\sum_i(\kappa_i'\mathbf{n}_i-\kappa_i^2\mathbf{t})\,\rangle =-2\sum_i(\kappa_i'\beta_i-\kappa_i^2\beta_0)\. \end{aligned}\] Imposing the order 3 contact condition leads to \(g'(s_0)=g''(s_0)=g'''(s_0)=0\) and gives \[\beta_0 = 0,\,\kappa_1(s_0)\beta _1+\kappa_2(s_0)\beta_2-1=0,\,\mbox{ and }\kappa_1'(s_0)\beta_1+\kappa_2'(s_0)\beta_2=0.\label{eq::coefOfOscSphUsingRMF}\] So, the coefficients \(\beta_0\), \(\beta_1\), and \(\beta_2\) as functions of \(s_0\) are \[\beta_0=0,\,\beta_1=\frac{\kappa_2'}{\kappa_1\kappa_2'-\kappa_1'\kappa_2}=\frac{\kappa_2'}{\tau\kappa^2},\,\mbox{ and }\beta_2=-\frac{\kappa_1'}{\kappa_1\kappa_2'-\kappa_1'\kappa_2}=-\frac{\kappa_1'}{\tau\kappa^2}\,,\] where in the equalities above we used the relation between \((\kappa_1,\kappa_2)\) and \((\kappa,\tau)\). *Proof of Theorem [\[thr::charSphCurves\]](#thr::charSphCurves){reference-type="ref" reference="thr::charSphCurves"} for \(C^4\) curves[^6].* Taking the derivative of the osculating center gives \[P_S' = \frac{{\rm d}}{{\rm d}s}\left(\alpha+\frac{\kappa_2'}{\tau\kappa^2}\mathbf{n}_1-\frac{\kappa_1'}{\tau\kappa^2}\mathbf{n}_2\right)\nonumber\\ =\left(\frac{{\rm d}}{{\rm d}s}\frac{\kappa_2'}{\tau\kappa^2}\right)\mathbf{n}_1-\left(\frac{{\rm d}}{{\rm d}s}\frac{\kappa_1'}{\tau\kappa^2}\right)\mathbf{n}_2\.\] From the linear independence of \(\{\mathbf{n}_1,\mathbf{n}_2\}\) we conclude that \(P'_S=0\), i.e., \(\alpha\) is spherical, if and only if \(\beta_1\) and \(\beta_2\) are constants. From Eq. ([\[eq::coefOfOscSphUsingRMF\]](#eq::coefOfOscSphUsingRMF){reference-type="ref" reference="eq::coefOfOscSphUsingRMF"}), this is equivalent to say that the normal development lies on a line not passing through the origin. 0◻
## Rotation minimizing frames for spherical curves
In order to find an RM frame along a curve we can compute the angle between the principal normal and an RM vector field. This can be done by conveniently using \(\theta'=\tau\), the only drawback being the need of a Frenet frame globally defined, i.e., no inflection point should be allowed, otherwise the torsion may be not well defined. This can be effectively done for spherical curves by extending a result in. Indeed, we have
*Proof.* (a) Assume \(p\) to be the origin (the general case is reduced to this one by studying \(\tilde{\alpha}=\alpha-p\)). The vectors \(\alpha/r,\,\alpha'\), and \((\alpha/r)\times\alpha'\) form an orthonormal frame along the curve[^7]. Write \[\alpha'' = \frac{1}{r}\langle\alpha'',\alpha\rangle\frac{\alpha}{r}+\langle\alpha'',\alpha'\rangle\,\alpha'+\frac{1}{r}\langle\alpha'',\alpha\times\alpha'\rangle\frac{\alpha}{r}\times\alpha'\.\] Since \(\alpha\) is parametrized by arc-length, we have \(\langle\alpha'',\alpha'\rangle=0\). In addition, from \(\langle\alpha,\alpha\rangle=r^2\), it follows that \(\langle\alpha'',\alpha\rangle=-\langle\alpha',\alpha'\rangle=-1\). In conclusion, the acceleration vector gives \[\alpha'' =-\frac{1}{r}\,\frac{\alpha}{r}+\frac{J}{r}\,\frac{\alpha}{r}\times\alpha'\,\Rightarrow\,\kappa=\Vert\alpha''\Vert=\frac{1}{r}\sqrt{1+J^2}\.\] Now, writing the normal and binormal vectors as \(\mathbf{n}=\alpha''/\kappa\) and \(\mathbf{b}=\alpha'\times\mathbf{n}=\alpha'\times\alpha''/\kappa\), and using the Frenet equation \(\tau=-\langle\mathbf{b}',\mathbf{n}\rangle\), we have \[\begin{aligned}
\tau & = &-\left\langle\frac{{\rm d}}{{\rm d}s}\left(\frac{\alpha'\times\alpha''}{\kappa}\right),\frac{\alpha''}{\kappa}\right\rangle\nonumber\\ & = &-\frac{1}{\kappa^2}\langle\alpha'\times\alpha''',\alpha''\rangle\nonumber\\ & = &-\frac{1}{r^2\kappa^2}\langle\alpha'\times\alpha''',-\alpha+J\,\alpha\times\alpha'\rangle. \end{aligned}\] Finally, using the expression for \(\kappa\) above, \(J'=\langle\alpha,\alpha'\times\alpha'''\rangle\), and the vector identity \(\langle \mathbf{A}\times\mathbf{B},\mathbf{C}\times\mathbf{D}\rangle = \langle \mathbf{A},\mathbf{C}\rangle\langle\mathbf{B},\mathbf{D}\rangle-\langle \mathbf{A},\mathbf{D}\rangle\langle\mathbf{B},\mathbf{C}\rangle\), we get the desired result for the torsion: \(\tau = J'/(1+J^2)\). (b) The expression for \(\theta\) is a consequence of the chain rule and of \((\arctan\,x)'=(1+x^2)^{-1}\). 0◻
# Behavior of a twisted curve near an osculating sphere
An alternative for the approach in the previous section is the observation that for a spherical curve \(\alpha\subset \mathbb{S}^2(p,r)\) the normals to the sphere along \(\alpha\), i.e., the normalized position vector \(\mathbf{N}=(\alpha-p)/r\), minimizes rotation. Indeed, \(\frac{{\rm d}}{{\rm d}s}(\alpha-p)/r=-(-1/r)\,\alpha'\) (this is an important step in the implementation of the *double reflection method* for computing approximations of RM frames ). The curvature \(\kappa_1\) associated to \(\mathbf{n}_1=(\alpha-p)/r\) is then \(\kappa_1=-1/r\). For the other RM vector field, i.e., \(\mathbf{n}_2=\alpha'\times(\alpha-p)/r\), one has \(\kappa_2=-J/r\). Using the concept of osculating spheres, we would intuitively say that every curve is locally spherical. In this case, it is tempting to ask if the normals to the osculating spheres minimize rotation. Unhappily, this strategy does not work unless the curve is spherical:
*Proof.* Using that \(R_S'=\rho'\varsigma/\tau R_S\) and \(P_S'=\varsigma\mathbf{b}\), direct computation of the derivative of \(\mathbf{N}=(\alpha-P_S)/R_S\) leads to Eq. ([\[eq::DerOfNormalToOscSphere\]](#eq::DerOfNormalToOscSphere){reference-type="ref" reference="eq::DerOfNormalToOscSphere"}). Finally, by a known result of geometry, the condition to be spherical leads to \(\varsigma\equiv0\), which by direct examination of Eq. ([\[eq::DerOfNormalToOscSphere\]](#eq::DerOfNormalToOscSphere){reference-type="ref" reference="eq::DerOfNormalToOscSphere"}) is a necessary and sufficient condition to have \(\mathbf{N}\) and \(\mathbf{t}\) parallel. 0◻Now we investigate how to extend Theorem [\[thr::KappaTauJforSpheres\]](#thr::KappaTauJforSpheres){reference-type="ref" reference="thr::KappaTauJforSpheres"} for a generic curve. Let \(\alpha\) be a regular twisted curve and \(\Sigma_s = \mathbb{S}^2(P_S(s),R_S(s))\) be its osculating sphere at \(\alpha(s)\). Near a fixed point \(\alpha(s_0)\) we can obtain a spherical curve \(\beta: (s_0-\epsilon,s_0+\epsilon)\to\Sigma_{s_0}\) by projecting \(\alpha\) on \(\Sigma_{s_0}\) according to \[\beta(t)=r_0\frac{\alpha(t)-a_0}{\Vert\alpha(t)-a_0\Vert}\,,\label{eq::OscSphericalProjection}\] where \(a_0=P_S(s_0)\) and \(r_0=R_S(s_0)\).
*Proof.* In order to compute \(\tau_{\beta}\) and \(\kappa_{\beta}\) it is enough to find \(\beta'\), \(\beta''\), and \(\beta'''\). Calculating the derivatives of \(\beta\) and taking into account the relations \[\left\{ \begin{array}{ccc} \langle\,\alpha(s_0)-a_0,\mathbf{t}(s_0)\,\rangle & = &0\\ 1+\kappa_{\alpha}(s_0)\,\langle\,\alpha(s_0)-a_0,\mathbf{n}(s_0)\,\rangle & = & 0\\ \kappa'_{\alpha}(s_0)-\kappa_{\alpha}^2(s_0)\tau_{\alpha}(s_0)\,\langle\,\alpha(s_0)-a_0,\mathbf{b}(s_0)\,\rangle & = & 0\\ \end{array} \right.\] satisfied by an osculating sphere, we obtain after some lengthy but straightforward calculations the following relations at \(s=s_0\) \[\left\{ \begin{array}{ccc} \beta'(s_0) &=& \mathbf{t}(s_0)\\ \beta''(s_0) &=& \kappa_{\alpha}(s_0)\mathbf{n}(s_0)\\ \beta'''(s_0)&=&-\kappa_{\alpha}^2(s_0)\mathbf{t}(s_0)+\kappa_{\alpha}'(s_0)\mathbf{n}(s_0)+\tau_{\alpha}(s_0)\kappa_{\alpha}(s_0)\mathbf{b}(s_0)\\ \end{array} \right.\,,\] where \(\{\mathbf{t},\mathbf{n},\mathbf{b}\}\) are the Frenet frame of \(\alpha\). It follows that \(\alpha(s_0)=\beta(s_0)\), \(\alpha'(s_0)=\beta'(s_0)\), \(\alpha''(s_0)=\beta''(s_0)\), and \(\alpha'''(s_0)=\beta'''(s_0)\) (this is not a surprise, since an osculating sphere has a contact of order 3 with its reference curve). Substituting the expressions above for \(\beta',\beta''\), and \(\beta''\), in the equation for the torsion and curvature and using the equalities \(\alpha^{(i)}(s_0)=\beta^{(i)}(s_0)\) (\(i=0,1,2,3\)) gives the desired result: \(\tau_{\beta}=\tau_{\alpha}\), \(\kappa_{\beta}=\kappa_{\alpha}\). Now let us express \(\kappa\) and \(\tau\) in terms of \(J\). First, observe that \(\beta\) is not necessarily parametrized by arc-length, so we must adapt the expressions in Theorem [\[thr::KappaTauJforSpheres\]](#thr::KappaTauJforSpheres){reference-type="ref" reference="thr::KappaTauJforSpheres"}. For a curve with a generic regular parameter \(t\), we can write \[\frac{{\rm d}\beta}{{\rm d}s}=\frac{1}{v}\frac{{\rm d}\beta}{{\rm d}t}\,\mbox{ and }\,\frac{{\rm d}^2\beta}{{\rm d}s^2}=-\frac{1}{v^3}\frac{{\rm d}v}{{\rm d}t}\frac{{\rm d}\beta}{{\rm d}t}+\frac{1}{v^2}\frac{{\rm d}^2\beta}{{\rm d}t^2}\,,\] where \(v(t)=\Vert\beta'(t)\Vert\). So, it follows that \[j(t)=j(s_{\beta}(t))=\frac{1}{v^3(t)}\langle\,\beta(t)-P_S(s_0),\,\beta'(t)\times\beta''(t)\,\rangle,\] where \(s_{\beta}\) is the arc-length parameter of \(\beta\) and \(j(s_{\beta})=\langle\,\beta(s_{\beta})-P_S(s_0),\,\beta'(s_{\beta})\times\beta''(s_{\beta})\,\rangle\). Finally, applying the expression from Theorem [\[thr::KappaTauJforSpheres\]](#thr::KappaTauJforSpheres){reference-type="ref" reference="thr::KappaTauJforSpheres"} to the spherical curve \(\beta\) and using that at \(t=s_0\) one has \(j(t=s_0)=J(s_0)\) and \(v(t=s_0)=1\), we find the expressions for \(\kappa_{\alpha}\) and \(\tau_{\alpha}\): in the second equality for \(\tau_{\alpha}\), we should use the fact that \(\textrm{d}P_S/\textrm{d}s=\varsigma\,\mathbf{b}\) in order to conclude that \(J'=\langle\alpha-P_S,\alpha'\times\alpha'''\rangle-\kappa\varsigma\). 0◻
# Characterization of plane curves via rotation minimizing frames
*Proof.* Since a plane curve \(\alpha\) lies on its osculating plane, then it lies on an RM moving plane: the principal normal vector of a plane curve is an RM vector. Conversely, let \(\alpha\) lies on an RM moving plane \(\mbox{span}\{\mathbf{t},\mathbf{n}_1\}\), i.e., \[\alpha(s)-p = A(s)\mathbf{t}(s)+B(s)\mathbf{n}_1(s),\] where \(p\) is constant. Taking the derivative gives \[\mathbf{t} = (A'-\kappa_1B)\mathbf{t}+(B'+\kappa_1A)\mathbf{n}_1+\kappa_2A\,\mathbf{n}_2\] and then \[\left\{ \begin{array}{ccc} A'-\kappa_1B & = & 1\\ B'+\kappa_1A & = & 0\\ \kappa_2A & = & 0\\ \end{array} \right.\.\label{eq::coefRMplaneCurve}\] If \(\kappa_2(s)=0\) for all \(s\), then \(\kappa=0\) or \(\tau=0\), since \(\kappa_2=\kappa\sin\theta\). In any case, the curve is planar. On the other hand, if \(A(s)=0\) for all \(s\), it follows from the second equation of ([\[eq::coefRMplaneCurve\]](#eq::coefRMplaneCurve){reference-type="ref" reference="eq::coefRMplaneCurve"}) that \(B\) is a constant. In this case, \(\alpha(s)-p = B\,\mathbf{n}_1(s)\) and the curve must be spherical: it lies on a sphere of radius \(\vert B\vert\) and center \(p\). In addition, from the first equation of ([\[eq::coefRMplaneCurve\]](#eq::coefRMplaneCurve){reference-type="ref" reference="eq::coefRMplaneCurve"}), it follows that \(\kappa_1=-B^{-1}\) is a constant. Now, using that for a spherical curve the normal development curve \((\kappa_1,\kappa_2)\) lies on a straight line not passing through the origin, Theorem [\[thr::charSphCurves\]](#thr::charSphCurves){reference-type="ref" reference="thr::charSphCurves"}, we find that \(\kappa_2\) is also a constant. Finally, since \(\kappa_1\) and \(\kappa_2\) are constants, we conclude that \(\kappa\) is a constant and \(\tau=0\), i.e., \(\alpha\) is a circle (of radius \(\vert B\vert\) and center \(p\)). 0◻
# Conclusion
In this work we studied characterizations of spherical and plane curves using RM frames. We first established a new proof for a known characterization of spherical curves by using osculating spheres described in terms of an RM frame. Going further with the study of spherical curves, we showed how to find the angle between the principal normal and an RM vector of a spherical curve. This was done by conveniently writing the curvature and torsion for a curve on a sphere. Later, we extended the study for the curvature and torsion for a generic curve using spherical analogs. Finally, we also addressed the problem of characterizing those curves whose position vector lies, up to a translation, on a plane spanned by the unit tangent and an RM vector and proved that they are precisely the plane curves. A problem that remains open is that of computing RM frames for a generic curve. To the best of our knowledge, no exact solution exists and in general one must resort to a numerical approach (e.g., the double reflection method ). We hope our investigations may give some hints in this direction.
[^1]: In fact, every curve is locally contained, up to second order, in its osculating plane. So, a plane curve must satisfy it globally.
[^2]: This follows from the fact that \(\langle \alpha-p,\alpha-p\rangle = r^2\) if and only if \(\langle \alpha',\alpha-p\rangle=0\).
[^3]: This is also valid for curves in the 3D sphere and hyperbolic space.
[^4]: For a level set surface \(\Sigma=G^{-1}(c)\), an order \(k\) contact is equivalent to \(\beta^{(i)}(s_0^*)=0\) (\(1\leq i\leq k\)), where \(\beta=G\circ\alpha\) and \(c=\beta(s_0^*)=\alpha(s_0)\).
[^5]: An attempt to extend these ideas in order to characterize curves that lie on a surface is described in.
[^6]: We need a \(C^4\) condition in order to compute \(\kappa_i''\): \(C^1\) is enough to have \(\mathbf{t}\); \(C^2\) to have \(\mathbf{t}'\) and then \(\kappa_i\); and \(C^3\) (\(C^4\)) to have \(\kappa_i'\) (\(\kappa_i''\)).
[^7]: Such frames are also known in the mathematical literature as Saban frames. They were introduced by Giacomo Saban in the characterization of spheres through the vanishing of \(\oint\rho^n\tau\,\mathrm{d}s\) for every closed curve on the surface, which generalizes the case for \(n=0\), i.e., \(\oint\tau\,\mathrm{d}s\), in Ref..
[^8]: For a generic analytic curve, the lateral limits \(\tau_0^-\) and \(\tau_0^+\) do exist and coincide. But if we drop the analyticity assumption, one of the lateral limit may diverge even for a \(C^{\infty}\) curve. Indeed, analytic curves are well behaved with respect to inflection points: given two analytic functions \(K\geq0\) and \(\tau\), there exists an analytic curve, up to a rigid motion of \(\mathbb{E}^3\), with curvature \(\kappa=\sqrt{K}\) and torsion \(\tau\). | {'timestamp': '2017-06-27T02:05:37', 'yymm': '1706', 'arxiv_id': '1706.01577', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01577'} |
# Introduction
Theory of isomonodromic deformation and Painlevé equations has long been a subject of intensive study, with many applications both in mathematics and in physics. Recent discovery of the isomonodromy/CFT correspondence , , opened up a new perspective to this domain. In a nutshell, it states that the Painlevé tau functions are Fourier transform of conformal blocks at \(c=1\). The AGT relation allows one to write a combinatorial formula for the latter, thereby giving an explicit series representation for generic Painlevé tau functions. The initial conjectures of , , concern the sixth Painlevé (PVI) equation, and at present three different proofs are known. In the first approach one constructs solutions to the linear problem out of conformal blocks with the insertion of a degenerate primary field. The second approach also makes use of CFT ideas, but leads directly to bilinear differential equations for tau functions. The third approach is based on the Riemann-Hilbert problem. It is shown that Fredholm determinant representation for tau functions reproduces the combinatorial series expansion, without recourse to conformal blocks or the AGT relation. Painlevé equations of other types are also studied by confluence, in the framework of irregular conformal blocks , through the AGT correspondence, and using Fredholm determinants. It is natural to ask whether a similar picture holds for the \(q\) difference analog of Painlevé equations. This has been studied in for the qPIII(\(D_8\)) equation. The aim of the present paper is to give a \(q\) analog of the construction of. Our method is based on the work, where the five-dimensional AGT correspondence was studied in the light of the quantum toroidal \(\mathfrak{gl}_1\) algebra (also known as the Ding-Iohara-Miki algebra). Having the central charge \(c=1\) corresponds to choosing \(t=q\) where \(t,q\) are the parameters of the algebra. This is a sort of classical limit, in the sense that the quantum algebra reduces to the enveloping algebra of a Lie algebra, see Section 2.1. We use the trivalent vertex operators of to introduce a \(q\) analog of primary fields and conformal blocks. As in the continuous case, a key property is the braiding relation in the presence of the degenerate field. Following the scheme of, we then write the solution \(Y(x)\) of the Riemann problem as Fourier transform of appropriate conformal block functions. For the sake of concreteness, we restrict our discussion to the setting relevant to the qPVI equation. Also we do not discuss the issue of convergence of these series. The above construction allows us to relate the unknown functions \(y,z\) of qPVI to tau functions. More specifically, we express \(y,z\) in terms of four tau functions which arise from the expansions at \(x=0\) or \(x=\infty\); see [\[y-tau\]](#y-tau){reference-type="eqref" reference="y-tau"}, [\[z-tau\]](#z-tau){reference-type="eqref" reference="z-tau"} below. On the other hand, in view of the work, we expect that there are four more tau functions which naturally enter the picture; these are the ones related to the singular points other than \(x=0,\infty\). However, for the lack of the notion of fusion, it is not clear to us how to related them to \(y,z\). The absence of fusion is the main obstacle in the \(q\) analog of isomonodromy theory. We present conjectural bilinear relations satisfied by these eight tau functions, see [\[bilin-1\]](#bilin-1){reference-type="eqref" reference="bilin-1"}--[\[bilin-8\]](#bilin-8){reference-type="eqref" reference="bilin-8"}. Assuming the conjecture we give the final expression for \(y,z\) in [\[yz-final\]](#yz-final){reference-type="eqref" reference="yz-final"}. The plan of the paper is as follows. In Section 2, we give a brief account of the results of, restricting to the special case \(t=q\). We introduce a \(q\) analog of primary fields, conformal blocks, degenerate \((2,1)\) fields and their braiding. In Section 3, we apply them to the \(q\) analog of the Riemann problem, focusing attention to the setting of the \(q\)PVI equation. We derive a formula expressing the unknowns \(y,z\) in terms of four tau functions. We then propose conjectural bilinear difference equations for eight tau functions. In Appendix we give a direct combinatorial proof of the braiding relation used in the text. *Notation.*Throughout the paper we fix \(q\in{\mathbb C}^\times\) such that \(|q|<1\). We set \([u]=(1-q^u)/(1-q)\), \((a;q)_N=\prod_{j=0}^{N-1}(1-aq^j)\), \((a_1,\ldots,a_k;q)_{\infty}=\prod_{j=1}^k(a_j;q)_\infty\), and \((a;q,q)_\infty=\prod_{j,k=0}^\infty(1-aq^{j+k})\). We use the \(q\) Gamma function, \(q\) Barnes function and the theta function \[\begin{aligned}
&\Gamma_q(u)=\frac{(q;q)_\infty}{(q^u;q)_\infty}(1-q)^{1-u}, \quad G_q(u)=\frac{(q^u;q,q)_\infty}{(q;q,q)_\infty}(q;q)_\infty^{u-1}(1-q)^{-(u-1)(u-2)/2}, \\ &\vartheta(u)=q^{u(u-1)/2}\Theta_q(q^u),\quad \Theta_q(x)=(x,q/x,q;q)_\infty, \end{aligned}\] which satisfy \(\Gamma_q(1)=G_q(1)=1\) and \[\begin{aligned}
&\Gamma_q(u+1)=[u]\Gamma_q(u), \quad G_q(u+1)=\Gamma_q(u)G_q(u), \\ &\vartheta(u+1)=-\vartheta(u)=\vartheta(-u)\. \end{aligned}\] A partition is a finite sequence of positive integers \(\lambda=(\lambda_1,\ldots,\lambda_l)\) such that \(\lambda_1\ge\ldots\ge\lambda_{l}>0\). We set \(\ell(\lambda)=l\). The conjugate partition \(\lambda'=(\lambda'_1,\ldots,\lambda'_{l'})\) is defined by \(\lambda'_j=\sharp\{i\mid \lambda_i\ge j\}\), \(l'=\lambda_1\). We denote by \(\Lambda\) the set of all partitions. We regard a partition \(\lambda\) also as the subset \(\{(i,j)\in{\mathbb Z}^2\mid 1\le j\le \lambda_i,\ i\ge 1\}\) of \({\mathbb Z}^2\), and denote its cardinality by \(|\lambda|\). For \(\square=(i,j)\in{\mathbb Z}_{>0}^2\) we set \(a_\lambda(\square)=\lambda_i-j\) and \(\ell_\lambda(\square)=\lambda'_j-i\). In the last formulas we set \(\lambda_i=0\) if \(i>\ell(\lambda)\) (resp. \(\lambda'_j=0\) if \(j>\ell(\lambda')\)).
# \(q\) analog of conformal blocks
In this section, we collect background materials from. We shall restrict ourselves to the special case \(t=q\), where formulas of simplify considerably. We then introduce a \(q\) analog of primary fields, conformal blocks and the braiding relation for the degenerate fields (analog of \((2,1)\) operators).
## Toroidal Lie algebra
Let \(Z,D\) be non-commuting variables satisfying \(ZD=qDZ\). The ring of Laurent polynomials \({\mathbb C}\langle Z^{\pm1},D^{\pm 1}\rangle\) in \(Z,D\) is a Lie algebra, with the Lie bracket given by the commutator. We consider its two-dimensional central extension, which we denote by \(\mathcal{L}\). As a vector space \(\mathcal{L}\) has \({\mathbb C}\)-basis \(Z^kD^l\) (\((k,l)\in{\mathbb Z}^2\setminus\{(0,0)\}\)) and central elements \(c_1,c_2\). The Lie bracket is defined by \[\begin{aligned}
&[Z^{k_1}D^{l_1},Z^{k_2}D^{l_2}]=(q^{-l_1k_2}-q^{-l_2k_1})Z^{k_1+k_2}D^{l_1+l_2} +\delta_{k_1+k_2,0}\delta_{l_1+l_2,0}q^{-k_2l_1} (k_1c_1+l_1 c_2)\. \end{aligned}\] We set \[\begin{aligned}
&a_r=Z^r,\quad x^+_n=DZ^n,\quad x^-_n=Z^nD^{-1}\,, \\ & b_r=D^{-r}, \quad y^+_n=Z D^{-n},\quad y^-_n=D^{-n}Z^{-1}\,, \end{aligned}\] and introduce the generating series \(x^\pm(z)=\sum_{n\in{\mathbb Z}} x^\pm_nz^{-n}\), \(y^\pm(z)=\sum_{n\in{\mathbb Z}} y^\pm_nz^{-n}\). We have two Heisenberg subalgebras of \(\mathcal{L}\), \(\mathfrak{a}=\oplus_{r\neq 0}{\mathbb C} a_r\oplus{\mathbb C} c_1\) and \(\mathfrak{b}=\oplus_{r\neq 0}{\mathbb C} b_r\oplus{\mathbb C} c_2\): \([a_r,a_s]=r\delta_{r+s,0}c_1\), \([b_r,b_s]=-r\delta_{r+s,0}c_2\). We say that an \(\mathcal{L}\)-module has level \((l_1,l_2)\) if \(c_1\) acts as \(l_1\cdot \textit{id}\) and \(c_2\) acts as \(-l_2\cdot \textit{id}\). The Lie algebra \(\mathcal{L}\) has automorphisms \(S,T\) given by \[\begin{aligned}
&S:Z\mapsto D,\quad D\mapsto Z^{-1},\quad c_1\mapsto c_2, \quad c_2\mapsto-c_1,\\ &T:Z\mapsto Z,\quad D\mapsto DZ,\quad c_1\mapsto c_1,\quad c_2\mapsto c_1+c_2\. \end{aligned}\] They satisfy \(S^4=(ST)^6=\textit{id}\), so that the group \(SL(2,{\mathbb Z})\) acts on \(\mathcal{L}\) by automorphisms. We have \(S(b_r)=a_r\), \(S(y^{\pm}(z))=x^\pm(z)\).
## Fock representations
The most basic representations of \(\mathcal{L}\) are the Fock representations of levels \((1,0)\) and \((0,1)\). Following we denote them by \({\mathcal F}^{(1,0)}_u\) and \({\mathcal F}^{(0,1)}_u\) (\(u\in {\mathbb C}^\times\)), respectively. The Fock representation \({\mathcal F}^{(1,0)}_u\) is irreducible under the Heisenberg subalgebra \(\mathfrak{a}\). As a vector space we have \({\mathcal F}^{(1,0)}_u={\mathbb C}[a_{-1},a_{-2},\ldots]\ket{\mathrm{vac}}\), where \(\ket{\mathrm{vac}}\) is a cyclic vector such that \(a_r\ket{\mathrm{vac}}=0\) (\(r>0\)). The action of the generators \(x^\pm_n\) is given by vertex operators \[\begin{aligned}
&x^\pm(z)\mapsto \frac{1}{1-q^{\mp1}}u^{\pm1} \exp\bigl(\mp\sum_{r\ge1}\frac{1-q^r}{r}a_{-r}z^r\bigr) \exp\bigl(\pm\sum_{r\ge1}\frac{1-q^{-r}}{r}a_{r}z^{-r}\bigr)\. \end{aligned}\] The Fock representation \({\mathcal F}^{(0,1)}_u\) is the pullback of \({\mathcal F}^{(1,0)}_u\) by the automorphism \(S\). It is irreducible with respect to the Heisenberg subalgebra \(\mathfrak{b}\), and \(y^\pm(z)\) act as vertex operators. On the other hand, the action of \(\mathfrak{a}\) on \({\mathcal F}^{(0,1)}_u\) is commutative. The joint eigenvectors \(\{\ket{\lambda}\}_{\lambda\in\Lambda}\) of \(\mathfrak{a}\) are labeled by all partitions and constitute a basis of \({\mathcal F}^{(0,1)}_u\). In particular, for the vector \(\ket{\emptyset}\) we have \[\begin{aligned}
a_r\ket{\emptyset}=-\frac{u^r}{1-q^r}\ket{\emptyset}\,, \quad x^-(z) \ket{\emptyset}=0, \quad x^+(z) \ket{\emptyset}=\delta(u/z)\ket{(1)} \end{aligned}\] where \(\delta(x)=\sum_{n\in{\mathbb Z}}x^n\). Formulas for the action of \(a_r\) and \(x^{\pm}(z)\) on a general \(\ket{\lambda}\) can also be written explicitly. Since we do not use them in this paper, we refer the reader to. Tensor products of \(N\) copies of \({\mathcal F}^{(1,0)}_u\) are closely related to representations of the \(W_N\) algebra (see e.g., ). We briefly explain this point in the case \(N=2\). Consider the tensor product of two Fock representations and its decomposition with respect to the Heisenberg subalgebra \(\mathfrak{a}\), \[\begin{aligned}
{\mathcal F}^{(1,0)}_{u_1}\otimes {\mathcal F}^{(1,0)}_{u_2}=\mathcal{H}\otimes \Omega_{u_1,u_2}\,, \end{aligned}\] where \(\mathcal{H}\) is the Fock space of \(\mathfrak{a}\) and \(\Omega_{u_1,u_2}\) is the multiplicity space. The operators \(x^{\pm}(z)\) act on \({\mathcal F}^{(1,0)}_{u_1}\otimes {\mathcal F}^{(1,0)}_{u_2}\) as a sum of two vertex operators. We can factor it into a product of two commuting operators, \[\begin{aligned}
(1-q^{\mp1})x^{\pm}(z)\mapsto g^\pm(z)\cdot (u_1u_2)^{\pm1/2}T(z;(u_1/u_2)^{\pm1/2}). \end{aligned}\] The first factor represents the action of \(\mathfrak{a}\) on \(\mathcal{H}\), \[\begin{aligned}
g^\pm(z)=:\exp\left(\pm\sum_{r\neq0}\frac{1-q^{-r}}{2r} \bigl(a^{(1)}_r+a^{(2)}_r\bigr)z^{-r}\right):\,, \end{aligned}\] where \(a_r^{(1)}=a_r\otimes \textit{id}\), \(a_r^{(2)}=\textit{id}\otimes a_r\). The second factor acts on \(\Omega_{u_1,u_2}\), \[\begin{aligned}
&T(z;u)=u \Lambda^+(z)+u^{-1}\Lambda^-(z),\\ &\Lambda^{\pm}(z)=:\exp\left(\pm\sum_{r\neq0}\frac{1-q^{-r}}{2r} \bigl(a^{(1)}_r-a^{(2)}_r\bigr)z^{-r}\right): \end{aligned}\] and gives a free field realization of the deformed Virasoro algebra. The space \(\Omega_{u_1,u_2}\) is a \(q\) analog of the Virasoro Verma module with central charge \(c=1\) and highest weight \(\Delta_\theta=\theta^2\), where \(q^{2\theta}=u_1/u_2\).
## Trivalent vertex operators
For \(N\in{\mathbb Z}\), we denote by \({\mathcal F}^{(1,N)}_w\) the pullback of \({\mathcal F}^{(1,0)}_w\) by the automorphism \(T^N\). The following intertwiners of \(\mathcal{L}\)-modules are called trivalent vertex operators. \[\begin{aligned}
&\Phi=\Phi\Bigl[{{(1,N+1),wx}\atop {(0,1),-w;(1,N),x}}\Bigr] :{\mathcal F}^{(0,1)}_{-w}\otimes {\mathcal F}^{(1,N)}_x \longrightarrow {\mathcal F}^{(1,N+1)}_{wx}\,, \\ &\Phi^*=\Phi^*\Bigl[{{(1,N),x;(0,1),-w}\atop {(1,N+1),wx}}\Bigr] :{\mathcal F}^{(1,N+1)}_{wx}\longrightarrow {\mathcal F}^{(1,N)}_x \otimes {\mathcal F}^{(0,1)}_{-w}\. \end{aligned}\] Such non-trivial intertwiners exist and are unique up to a scalar multiple. We introduce their components \(\widehat{\Phi}_\lambda(w),\widehat{\Phi}^*_\lambda(w)\) relative to the basis \(\{\ket{\lambda}\}\) of \({\mathcal F}^{(0,1)}_{-w}\) as follows. \[\begin{aligned}
&\Phi\bigl(\ket{\lambda}\otimes \alpha\bigr)= \hat{t}(\lambda,x,w,N) \cdot \widehat{\Phi}_\lambda(w)(\alpha)\quad (\alpha\in {\mathcal F}^{(1,N)}_x)\,, \\ &\Phi^* (\alpha) =\sum_{\lambda\in\Lambda} \hat{t}^*(\lambda,x,w,N) \cdot \widehat{\Phi}^*_\lambda(w)(\alpha)\otimes \ket{\lambda} \quad (\alpha\in {\mathcal F}^{(1,N+1)}_{wx}) \. \end{aligned}\] The numerical factors in front are \[\begin{aligned}
\hat{t}(\lambda,x,w,N) =\frac{q^{n(\lambda')}}{c_\lambda}f_\lambda^{-N-1} \left(\frac{x}{w^N}\right)^{|\lambda|}\,, \quad \hat{t}^*(\lambda,x,w,N) =\frac{q^{n(\lambda')}}{c_\lambda}f_\lambda^N \left(\frac{q w^N}{x}\right)^{|\lambda|} \,, \end{aligned}\] where \[\begin{aligned}
&n(\lambda) =\sum_{\square\in\lambda}\ell_\lambda(\square),\quad n(\lambda') =\sum_{\square\in\lambda}a_\lambda(\square), \\ &f_\lambda=(-1)^{|\lambda|}q^{n(\lambda')-n(\lambda)}\,, \quad c_\lambda=\prod_{\square\in\lambda}(1-q^{\ell_\lambda(\square)+a_\lambda(\square)+1})\. \end{aligned}\] We normalize \(\widehat{\Phi}(w)\), \(\widehat{\Phi}^*(w)\) so that \(\widehat{\Phi}_\emptyset(w)\ket{\mathrm{vac}}=\ket{\mathrm{vac}}+\cdots\), \(\widehat{\Phi}^*_\emptyset(w)\ket{\mathrm{vac}}=\ket{\mathrm{vac}}+\cdots\).
Of particular importance is their normal ordering rule. For a pair of partitions \((\lambda,\mu)\in\Lambda^2\), we introduce the Nekrasov factor by \[\begin{aligned}
&N_{\lambda,\mu}(w) =\prod_{\square\in\lambda}(1-q^{-\ell_\lambda(\square)-a_{\mu}(\square)-1}w) \prod_{\square\in\mu}(1-q^{a_{\lambda}(\square)+\ell_\mu(\square)+1}w) \. \end{aligned}\]
It is useful to note the relations \[\begin{aligned}
&N_{\lambda,\mu}(w)=N_{\mu,\lambda}(w^{-1}) w^{|\lambda|+|\mu|}\frac{f_\lambda}{f_\mu}, \label{rule1}\\ &\left(\frac{q^{n(\lambda')}}{c_\lambda}\right)^2=\frac{f_\lambda q^{-|\lambda|}}{N_{\lambda,\lambda}(1)}\. \label{rule2} \end{aligned}\]
## Chiral primary field
In what follows we set \[\begin{aligned}
{\mathcal V}_{\theta,w}:={\mathcal F}^{(0,1)}_{-q^{-\theta}w}\otimes {\mathcal F}^{(0,1)}_{-q^{\theta}w}\. \end{aligned}\] We fix its basis \(\ket{\boldsymbol{\lambda}}=\ket{\lambda_+}\otimes\ket{\lambda_-}\) labeled by a pair \(\boldsymbol{\lambda}=(\lambda_+,\lambda_-)\) of partitions. Consider the compositions of trivalent vertex operators \[\begin{aligned}
&\Psi_1:{\mathcal F}^{(1,0)}_{q^{-2\theta_2}/x} \longrightarrow {\mathcal F}^{(1,-1)}_{q^{-\theta_3-\theta_2}/(xw)}\otimes {\mathcal F}^{(0,1)}_{-q^{\theta_3-\theta_2}w} \longrightarrow {\mathcal F}^{(1,-2)}_{1/(xw^2)}\otimes {\mathcal F}^{(0,1)}_{-q^{-\theta_3-\theta_2}w} \otimes {\mathcal F}^{(0,1)}_{-q^{\theta_3-\theta_2}w}\,, \\ &\Psi_2: {\mathcal F}^{(0,1)}_{-q^{-\theta_1}w}\otimes {\mathcal F}^{(0,1)}_{-q^{\theta_1}w}\otimes {\mathcal F}^{(1,-2)}_{1/(xw^2)} \longrightarrow {\mathcal F}^{(0,1)}_{-q^{-\theta_1}w}\otimes {\mathcal F}^{(1,-1)}_{q^{\theta_1}/(xw)} \longrightarrow {\mathcal F}^{(1,0)}_{1/x}\. \end{aligned}\] By composing them we obtain an intertwiner of \(\mathcal{L}\)-modules \[\begin{aligned}
{\mathcal V}_{\theta_1,w}\otimes {\mathcal F}^{(1,0)}_{q^{-2\theta_2}/x} \overset{\textit{id}\otimes\Psi_1}{\longrightarrow} {\mathcal V}_{\theta_1,w}\otimes {\mathcal F}^{(1,-2)}_{1/(xw^2)} \otimes {\mathcal V}_{\theta_3,q^{-\theta_2}w} \overset{\Psi_2\otimes\textit{id}}{\longrightarrow} {\mathcal F}^{(1,0)}_{1/x}\otimes{\mathcal V}_{\theta_3,q^{-\theta_2}w}\. \label{comp} \end{aligned}\] Taking further the vacuum-to-vacuum matrix coefficient of [\[comp\]](#comp){reference-type="eqref" reference="comp"}, we obtain a map \[\begin{aligned}
&V\Bigl({{\phantom{h_1} \theta_2 \phantom{h_3}} \atop{\theta_3 \phantom{h_2} \theta_1}};w,x\Bigr) :{\mathcal V}_{\theta_1,w}\longrightarrow {\mathcal V}_{\theta_3,q^{-\theta_2}w}\. \label{primary} \end{aligned}\] The matrix coefficients of [\[primary\]](#primary){reference-type="eqref" reference="primary"} are given by \[\begin{aligned}
&\langle{\boldsymbol{\mu}}\Bigl| V\Bigl({{\phantom{h_1} \theta_2 \phantom{h_3}} \atop{\theta_3 \phantom{h_2} \theta_1}};w,x\Bigr) \Bigr|\boldsymbol{\lambda}\rangle \label{V-matelt}\\ &=\mathcal{C}\, \frac{q^{n(\lambda'_+)}}{c_{\lambda_+}}\frac{q^{n(\lambda'_-)}}{c_{\lambda_-}} \frac{q^{n(\mu'_+)}}{c_{\mu_+}}\frac{q^{n(\mu'_-)}}{c_{\mu_-}} \frac{f_{\lambda_-}}{f_{\mu_+}^2f_{\mu_-}}q^{2\theta_1|\lambda_-| +(2(\theta_3+\theta_2)+1)|\mu_+|+(2\theta_2+1)|\mu_-|}\, \nonumber\\ &\times x^{|\boldsymbol{\mu}|-|\boldsymbol{\lambda}|}\cdot \bra{\mathrm{vac}} \widehat{\Phi}_{\lambda_+}(q^{-\theta_1}w) \widehat{\Phi}_{\lambda_-}(q^{\theta_1}w) \widehat{\Phi}^*_{\mu_+}(q^{-\theta_3-\theta_2}w) \widehat{\Phi}^*_{\mu_-}(q^{\theta_3-\theta_2}w)\ket{\mathrm{vac}}\,, \nonumber \end{aligned}\] where we set \(|\boldsymbol{\lambda}|=|\lambda_+|+|\lambda_-|\) for \(\boldsymbol{\lambda}=(\lambda_+,\lambda_-)\in\Lambda^2\), and \(\mathcal{C}\) is a scalar factor independent of \(\boldsymbol{\lambda},\boldsymbol{\mu}\in\Lambda^2\). We choose the normalization \[\begin{aligned}
&\langle V\Bigl({{\phantom{h_3} \theta_2 \phantom{h_1}} \atop{\theta_3 \phantom{h_2} \theta_1}};w,x\Bigr) \rangle = {\mathcal N} \Bigl({{\phantom{h_3} \theta_2 \phantom{h_1}} \atop{\theta_3 \phantom{h_2} \theta_1}}\Bigr) q^{2\theta_2\theta_3^2}x^{\theta_3^2-\theta_2^2-\theta_1^2}\,, \nonumber \\ &{\mathcal N} \Bigl({{\phantom{h_3} \theta_2 \phantom{h_1}} \atop{\theta_3 \phantom{h_2} \theta_1}}\Bigr) = \frac{\prod_{\epsilon,\epsilon'=\pm} G_q(1+\epsilon\theta_3-\theta_2-\epsilon'\theta_1)} {G_q(1+2\theta_3)G_q(1-2\theta_1)}\. \label{Nfac} \end{aligned}\] Here and after, \(\langle\cdots\rangle\) stands for the matrix coefficient \(\langle(\emptyset,\emptyset)|\cdots|(\emptyset,\emptyset\rangle)\). We regard [\[primary\]](#primary){reference-type="eqref" reference="primary"} as a \(q\) analog of the chiral primary field of conformal dimension \(\Delta_{\theta_2}=\theta_2^2\).
## Conformal block function
The \((m+2)\) point conformal block function is the expectation value of a product of chiral primary fields, \[\begin{aligned}
&{\mathcal F}\Bigl( { {\phantom{000} \theta_m\phantom{h000} \theta_{m-1}\phantom{h0}\phantom{h0} \cdots\phantom{h0} \phantom{h0}\theta_1\phantom{h0}} \atop {\theta_{m+1}\phantom{h0}\sigma_{m-1} \phantom{h0}\sigma_{m-2}\phantom{h0} \cdots \phantom{h_0} \sigma_{1}\phantom{h0}\theta_{0}} } ;x_m,\ldots,x_{1} \Bigr) \\ &=\langle V\Bigl({\kern-1pt\theta_m \atop{\theta_{m+1} \phantom{h_2} \sigma_{m-1}}};w_m,x_m\Bigr) V\Bigl({{\theta_{m-1}\phantom{h}} \atop{\sigma_{m-1} \phantom{h_2} \sigma_{m-2}}};w_{m-1},x_{m-1}\Bigr) \cdots V\Bigl({{\theta_1 \phantom{h}} \atop{\sigma_1 \phantom{h} \theta_{0}}};w_{1},x_{1}\Bigr) \rangle \end{aligned}\] where \(w_p=q^{-\theta_{p-1}-\cdots-\theta_1}w_1\). The right hand side is actually independent of \(w_1\). Using formula [\[V-matelt\]](#V-matelt){reference-type="eqref" reference="V-matelt"}, applying the normal ordering rule [\[normal\]](#normal){reference-type="eqref" reference="normal"} and further using [\[rule1\]](#rule1){reference-type="eqref" reference="rule1"}, [\[rule2\]](#rule2){reference-type="eqref" reference="rule2"} for simplification, we obtain the following explicit expression. \[\begin{aligned}
&{\mathcal F}\Bigl( { {\phantom{000} \theta_m\phantom{h000} \theta_{m-1}\phantom{h0}\phantom{h0} \cdots\phantom{h0} \phantom{h0}\theta_1\phantom{h0}} \atop {\theta_{m+1}\phantom{h0}\sigma_{m-1} \phantom{h0}\sigma_{m-2}\phantom{h0} \cdots \phantom{h_0} \sigma_{1}\phantom{h0}\theta_{0}} } ;x_m,\ldots,x_{1} \Bigr) \label{block}\\ &= \prod_{p=1}^{m}{\mathcal N}\Bigl({\kern-1pt\theta_p\atop{\sigma_{p} \phantom{h_2} \sigma_{p-1}}} \Bigr)q^{2\theta_p\sigma_{p}^2} \cdot \prod_{p=1}^{m} x_p^{\sigma_{p}^2-\theta_p^2-\sigma_{p-1}^2} \nonumber\\ &\times \sum_{\boldsymbol{\lambda}^{(1)},\ldots,\boldsymbol{\lambda}^{(m-1)}} \prod_{p=1}^{m-1}\Bigl( \frac{q^{2\theta_{p}}x_p}{x_{p+1}} \Bigr)^{|\boldsymbol{\lambda}^{(p)}|}\cdot \frac{\prod_{p=1}^m \prod_{\epsilon,\epsilon'=\pm} N_{\lambda^{(p)}_\epsilon,\lambda^{(p-1)}_{\epsilon'}}(q^{\epsilon \sigma_{p}-\theta_p-\epsilon'\sigma_{p-1}})} {\prod_{p=1}^{m-1}\prod_{\epsilon,\epsilon'=\pm} N_{\lambda^{(p)}_{\epsilon},\lambda^{(p)}_{\epsilon'}}(q^{\epsilon\sigma_p-\epsilon'\sigma_p}) }\. \nonumber \end{aligned}\] We have set \(\sigma_0=\theta_0\), \(\sigma_m=\theta_{m+1}\), \(\boldsymbol{\lambda}^{(0)}=\boldsymbol{\lambda}^{(m)}=(\emptyset,\emptyset)\), and the sum is taken over all \(\boldsymbol{\lambda}^{(p)}=(\lambda^{(p)}_+,\lambda^{(p)}_-)\in\Lambda^2\), \(p=1,\ldots,m-1\). Formula [\[block\]](#block){reference-type="eqref" reference="block"} is the \(q\) version of the AGT formula for conformal blocks specialized to the case \(t=q\). The derivation given above from the quantum algebra viewpoint is due to.
## Degenerate field and braiding
The normalization factor [\[Nfac\]](#Nfac){reference-type="eqref" reference="Nfac"} has a zero when \(\theta_2=1/2\) and \(\theta_1-\theta_3=\pm 1/2\). In this case we redefine the primary field by \[\begin{aligned}
{\mathcal N}'\Bigl({{\! \textstyle{\frac 1 2} \phantom{h_3}} \atop{\theta \phantom{h} \ \ \theta\pm \textstyle{\frac 1 2}}}\Bigr) = \lim_{\varepsilon\to 0}\,\frac{1}{G_q(\varepsilon)}\, {\mathcal N}\Bigl({{\phantom{h} \textstyle{\frac 1 2}-\varepsilon \phantom{hhh}} \atop{\theta \phantom{h_2} \ \ \theta\pm \textstyle{\frac 1 2}}}\Bigr)\,, \\ V'\Bigl({{\phantom{h} \textstyle{\frac 1 2} \phantom{hhh}} \atop{\theta \phantom{h_2} \theta\pm \textstyle{\frac 1 2}}};w,x\Bigr) = \lim_{\varepsilon\to 0}\,\frac{1}{G_q(\varepsilon)}\, V\Bigl({{\phantom{h} \textstyle{\frac 1 2}-\varepsilon \phantom{hhh}} \atop{\theta \phantom{h_2} \ \ \theta\pm \textstyle{\frac 1 2}}};w,x\Bigr)\. \end{aligned}\] We consider the conformal block function when one of the primary fields is \(V'\Bigl(\cdots\Bigr)\). We keep using the same letter \({\mathcal F}\left(\cdots\right)\) to denote them. By choosing \(m=2\) and specializing parameters in [\[block\]](#block){reference-type="eqref" reference="block"}, the four point conformal blocks can be evaluated in terms of Heine's basic hypergeometric series \[\begin{aligned}
F\left({{\alpha,\beta}\atop{\gamma}};x\right) =\frac{\Gamma_q(\alpha)\Gamma_q(\beta)}{\Gamma_q(\gamma)} {}_2\kern-1pt\phi_1\left({{q^\alpha,q^\beta}\atop{q^\gamma}};q,x\right)\, \end{aligned}\] as follows. \[\begin{aligned}
&{\mathcal F}\Bigl({{\kern-3pt \textstyle{\frac 1 2}\phantom{hhhhh} \theta_1\phantom{hh}} \atop {\theta_\infty\phantom{h0}\theta_\infty+\textstyle{\frac{\varepsilon'}{2}} \phantom{h0}\theta_0}} ;x_1,x_2\Bigr) = \mathcal{N}\, q^{\theta_\infty^2}\left(\frac{q^{2\theta_1}x_2}{x_1}\right)^{\varepsilon'\theta_\infty+1/4} \\ &\qquad \times F\left({{\textstyle{\frac 1 2}+\varepsilon'\theta_\infty-\theta_1+\theta_0, \textstyle{\frac 1 2}+\varepsilon'\theta_\infty-\theta_1-\theta_0} \atop{1+2\varepsilon'\theta_\infty}};\frac{q^{2\theta_1}x_2}{x_1}\right) \,, \\ &{\mathcal F}\Bigl({{ \theta_1\phantom{hhh} \textstyle{\frac 1 2} \phantom{h}} \atop {\theta_\infty\phantom{h0}\theta_0+\textstyle{\frac{\varepsilon}{2}} \phantom{h0}\theta_0}} ;x_2,x_1\Bigr) =\mathcal{N}\, q^{\theta_0^2}\left(\frac{qx_1}{x_2}\right)^{\varepsilon\theta_0+1/4} \\ &\qquad\times F\left({{\textstyle{\frac 1 2}+\theta_\infty-\theta_1+\varepsilon\theta_0, \textstyle{\frac 1 2}-\theta_\infty-\theta_1+\varepsilon\theta_0} \atop{1+2\varepsilon\theta_0}};\frac{qx_1}{x_2}\right) \. \end{aligned}\] Here we have set \[\begin{aligned}
\mathcal{N}=q^{2\theta_1\theta_\infty^2} x_1^{-1/4}x_2^{\theta_\infty^2-\theta_1^2-\theta_0^2} \frac{\prod_{\mu,\mu'=\pm} G_q\Bigl(\frac{1}{2}+\mu\theta_\infty-\theta_1+\mu'\theta_0\Bigr)} {G_q(1+2\theta_\infty)G_q(1-2\theta_0)}. \end{aligned}\] The following braiding relation holds.
For the vacuum matrix element, the relation [\[braiding-rel\]](#braiding-rel){reference-type="eqref" reference="braiding-rel"} is a consequence of the known connection formula for basic hypergeometric functions. At this writing it is not clear to us whether the general case can be deduced from this and the intertwining relation. In Appendix we give a direct combinatorial proof of the braiding relation [\[braiding-rel\]](#braiding-rel){reference-type="eqref" reference="braiding-rel"}. The braiding matrix [\[braid\]](#braid){reference-type="eqref" reference="braid"} has the following properties. \[\begin{aligned}
&{\mathcal B}_{\varepsilon,\varepsilon'}\left[\begin{matrix} \theta_1 & \textstyle{\frac 1 2} \\ \theta_\infty & \theta_0\\ \end{matrix} \Bigl| q x\right] ={\mathcal B}_{\varepsilon,\varepsilon'}\left[\begin{matrix} \theta_1 & \textstyle{\frac 1 2} \\ \theta_\infty & \theta_0\\ \end{matrix} \Bigl| x\right]\,, \label{braid-q-per}\\ &{\mathcal B}\left[\begin{matrix} \theta_1 & \textstyle{\frac 1 2} \\ \theta_\infty & \theta_0\\ \end{matrix} \Bigl| x\right]^{-1} = {\mathcal B}\left[\begin{matrix} \theta_1 & \textstyle{\frac 1 2} \\ \theta_0 & \theta_\infty\\ \end{matrix} \Bigl| q^{-2\theta_1}x^{-1}\right], \label{braid-inv} \\ & \det {\mathcal B}\left[\begin{matrix} \theta_1 & \textstyle{\frac 1 2} \\ \theta_\infty & \theta_0\\ \end{matrix} \Bigl| x\right] =\frac{\vartheta\bigl(2\theta_\infty\bigr)}{\vartheta\bigl(2\theta_0\bigr)} \frac{\vartheta\bigl(u\bigr)}{\vartheta\bigl(u+2\theta_1\bigr)}\,, \nonumber \\ &{\mathcal B}\left[ \begin{matrix} \theta_1 & \textstyle{\frac 1 2} \\ \theta_\infty & \theta_0\\ \end{matrix} \Bigl| x\right] \quad \text{is \(1\)-periodic in \(\theta_0\),\(\theta_1\),\(\theta_\infty\)}\. \label{braid-period} \end{aligned}\]
# \(q\)-Painlevé VI equation
## Riemann problem
This paper deals with a nonlinear nonautonomous \(q\)-difference equation called \(q\)-Painlevé VI (\(q\)PVI) equation. It is derived from deformation theory of linear \(q\)-difference equations as a discrete analog of isomonodromic deformation. A remark is in order concerning the name of the equation. Discrete Painlevé equations are classified by rational surface theory and sometimes called by the name of the surface. From that viewpoint, \(q\)PVI is an equation corresponding to the surface of type \(A_3^{(1)}\). Besides, this surface has symmetry by the Weyl group \(D_5^{(1)}\). Accordingly \(q\)PVI is also referred to as the \(q\)-Painlevé equation of type \(A_3^{(1)}\), or the \(q\)-Painlevé equation with \(D_5^{(1)}\)-symmetry. Regarding discrete Painlevé equations, pioneering researches by Ramani, Grammaticos, and their coworkers are well-known. The \(q\)PVI equation, in the special case when an extra symmetry condition is fulfilled, has been studied earlier by them under the name of discrete Painlevé III equation. In this section we review the content of concerning the generalized Riemann problem associated with the \(q\)PVI equation. Let \(\theta_0,\theta_t,\theta_1,\theta_\infty\in{\mathbb C}\). We set \[\begin{aligned}
&R_1=\max(1,|q^{-2\theta_1}|,|tq^{-2\theta_1}|,|tq^{-2\theta_1-2\theta_t}|), \\ &R_2=\min(|q^{-1}|,|q^{-2\theta_1-1}|, |tq^{-2\theta_1-1}|,|tq^{-2\theta_1-2\theta_t-1}|), \end{aligned}\] and assume that \(R_1<R_2\). Fix also \(\sigma\in{\mathbb C}\) and \(s\in{\mathbb C}^\times\). Suppose we are given \(2\times 2\) matrices \(Y^\infty(x,t),Y^{0t}(x,t),Y^0(x,t)\) with the following properties.
1. These matrices are holomorphic in the domains \[\begin{aligned}
Y^\infty(x,t): R_1<|x|\,, \quad Y^{0t}(x,t): R_1<|x|<R_2\,, \quad Y^0(x,t): |x|<R_2\. \label{domain} \end{aligned}\]
2. They are related to each other by \[\begin{aligned}
&Y^\infty(x,t)=Y^{0t}(x,t){\mathcal B}_1(x), \quad Y^{0t}(x,t)=Y^0(x,t){\mathcal B}_2(x), \label{connection1} \\ &{\mathcal B}_1(x) ={\mathcal B}\left[\begin{matrix} \theta_1 &\textstyle{\frac 1 2}\\ \theta_\infty+\textstyle{\frac 1 2} & \sigma \\ \end{matrix} \Bigl|q^{-2\theta_1} x^{-1} \right]\,, \quad {\mathcal B}_2(x) = {\mathcal B}\left[\begin{matrix} \theta_t& \textstyle{\frac 1 2} \\ \sigma+\textstyle{\frac 1 2} & \theta_0\\ \end{matrix} \Bigl|q^{-2\theta_t-2\theta_1}\frac{t}{x}\right] \begin{pmatrix} 0 & s \\ 1 & 0 \\ \end{pmatrix} \,, \label{connection2} \end{aligned}\]
3. At \(x\to \infty\) or \(x\to 0\) they have the behavior \[\begin{aligned}
&Y^\infty(x,t)=\left(I+Y_1(t) x^{-1}+O(x^{-2})\right) \begin{pmatrix} x^{-\theta_\infty} & 0 \\ 0 & x^{\theta_\infty}\\ \end{pmatrix} \quad (x\to \infty), \label{Yinf}\\ &Y^0(x,t)=G(t)(I+O(x)) \begin{pmatrix} x^{\theta_0} & 0 \\ 0 & x^{-\theta_0}\\ \end{pmatrix}\times x^{-\theta_t-\theta_1} \quad (x\to 0). \label{Y0} \end{aligned}\]
It follows that \(Y(x,t)=Y^\infty(x,t)\) is a solution of linear \(q\)-difference equations of the form \[\begin{aligned}
&Y(qx,t)=A(x,t)Y(x,t)\,,\quad A(x,t)=\frac{A_2 x^2+A_1(t)x+A_0(t)}{(x-q^{-1})(x-tq^{-2\theta_1-1})}\,, \label{A} \\ &Y(x,qt)=B(x,t)Y(x,t)\,, \quad B(x,t)=\frac{xI+B_0(t)}{x-tq^{-2\theta_t-2\theta_1}}\. \label{B} \end{aligned}\] We have \[\begin{aligned}
&A_2=\mathrm{diag}(q^{-\theta_\infty},q^{\theta_\infty}), \quad A_0(t)=tq^{-3\theta_1-\theta_t-2} G(t)\cdot \mathrm{diag}(q^{\theta_0},q^{-\theta_0}) \cdot G(t)^{-1}, \\ &\det A(x,t)=\frac{(x-q^{-2\theta_1-1})(x-t q^{-2\theta_t-2\theta_1-1})} {(x-q^{-1})(x-tq^{-2\theta_1-1})}\. \end{aligned}\] Though the argument is quite standard, we outline the derivation for completeness. We take the determinant of the relations [\[connection1\]](#connection1){reference-type="eqref" reference="connection1"}, [\[connection2\]](#connection2){reference-type="eqref" reference="connection2"} and rewrite them as \[\begin{aligned}
& \frac{(1/x,tq^{-2\theta_1}/x;q)_\infty} {(q^{-2\theta_1}/x,tq^{-2\theta_t-2\theta_1}/x;q)_\infty} \det Y^{\infty}(x,t)\, \\ &=-\frac{\vartheta(2\theta_\infty)}{\vartheta(2\sigma)} \frac{(q^{2\theta_1+1}x,q^{-2\theta_1}t/x;q)_\infty} {(qx,tq^{-2\theta_t-2\theta_1}/x;q)_\infty} q^{2\theta_1^2+\theta_1}x^{2\theta_1} \det Y^{0t}(x,t) \\ &=-s\frac{\vartheta(2\theta_\infty)}{\vartheta(2\theta_0)} \frac{(q^{2\theta_1+1}x,q^{2\theta_t+2\theta_1+1}x/t;q)_\infty} {(qx,q^{2\theta_1+1}x/t;q)_\infty} q^{2\theta_t^2+\theta_t+2\theta_1^2+\theta_1} \left(q^{2\theta_1}\frac{x}{t}\right)^{2\theta_t}x^{2\theta_1} \det Y^0(x,t)\. \end{aligned}\] Both sides of this expression are single-valued and holomorphic on \({\mathbb C}^\times\). In view of the behavior [\[Yinf\]](#Yinf){reference-type="eqref" reference="Yinf"}, [\[Y0\]](#Y0){reference-type="eqref" reference="Y0"} we conclude that \[\begin{aligned}
\det Y^{\infty}(x,t) = \frac{(q^{-2\theta_1}/x,tq^{-2\theta_t-2\theta_1}/x;q)_\infty} {(1/x,tq^{-2\theta_1}/x;q)_\infty}\. \end{aligned}\] In particular, the matrices \(Y^\infty(x,t)\), \(Y^{0t}(x,t)\) and \(Y^0(x,t)\) are generically invertible. Rewriting [\[connection1\]](#connection1){reference-type="eqref" reference="connection1"}, [\[connection2\]](#connection2){reference-type="eqref" reference="connection2"} and using the periodicity of the braiding matrices, we obtain \[\begin{aligned}
A(x,t):=Y^\infty(qx,t){Y^\infty(x,t)}^{-1} =Y^{0t}(qx,t){Y^{0t}(x,t)}^{-1} =Y^0(qx,t){Y^0(x,t)}^{-1}\. \end{aligned}\] A similar argument shows that \(A(x,t)\) is a rational function of \(x\) with the only poles at \(x=q^{-1},q^{-2\theta_1-1}t\). Using [\[Yinf\]](#Yinf){reference-type="eqref" reference="Yinf"}, [\[Y0\]](#Y0){reference-type="eqref" reference="Y0"} once again we find that \(A(x,t)\) has the form stated above. The derivation of [\[B\]](#B){reference-type="eqref" reference="B"} is quite similar. Introduce \(y=y(t)\), \(z=z(t)\) and \(w=w(t)\) by \[\begin{aligned}
&A(x,t)_{+-} =\frac{ q^{\theta_\infty}w(x-y) } {(x-q^{-1})(x-tq^{-2\theta_1-1})}\,, \\ &A(y,t)_{++}=\frac{y-tq^{-2\theta_t-2\theta_1-1}}{qz(y-q^{-1})} \,, \end{aligned}\] where we use \(\pm\) to indicate components of a \(2\times2\) matrix. Notation being as above, the compatibility \[A(x,qt)B(x,t)=B(qx,t)A(x,t)\label{compatibility}\] of [\[A\]](#A){reference-type="eqref" reference="A"} and [\[B\]](#B){reference-type="eqref" reference="B"} leads to the \(q\)PVI equation. Here and after we write \(\bar{f}(t)=f(qt)\), \(\underline{f}(t)=f(q^{-1}t)\).
## CFT construction
We are now in a position to construct solutions of the generalized Riemann problem in terms of conformal block functions. Following the method of, we consider the following sums of 5 point conformal blocks, \[\begin{aligned}
&Y^{\infty}_{\varepsilon,\varepsilon'}(x,t)= \frac{1}{k_\infty(\varepsilon)} \sum_{n\in{\mathbb Z}}s^n {\mathcal F}\Bigl({{\phantom{hh00}\textstyle{\frac 1 2} \phantom{hhhh0000} \theta_1\phantom{hhhhh} \theta_t \phantom{hhhhh}} \atop {\theta_\infty-\textstyle{\frac{\varepsilon}{2}} \phantom{h0}\theta_\infty-\textstyle{\frac{\varepsilon-\varepsilon'}{2}} \phantom{h0}\sigma+n \phantom{h0} \theta_0 }} ;q^{2\theta_t+2\theta_1}x,q^{2\theta_t},t\Bigr) \,, \label{CFT-Yinf}\\ &Y^{0t}_{\varepsilon,\varepsilon'}(x,t) =\frac{x^{-\theta_1}}{k_{0t}(\varepsilon)} \sum_{n\in{\mathbb Z}}s^n {\mathcal F}\Bigl({{\phantom{0000}\theta_1 \phantom{hhhhh00000}\textstyle{\frac 1 2} \phantom{hhhhh} \theta_t \phantom{hhhhh00}} \atop {\theta_\infty-\textstyle{\frac{\varepsilon}{2}} \phantom{h0}\sigma+n+\textstyle{\frac{\varepsilon'}{2}} \phantom{h0}\sigma+n \phantom{h0}\theta_0 } } ;q^{2\theta_t},q^{2\theta_t+2\theta_1}x,t\Bigr) \label{CFT-Y0t}\\ &\qquad \qquad =\frac{x^{-\theta_1}}{k_{0t}(\varepsilon)} \sum_{n\in{\mathbb Z}}s^{n+\frac{1-\varepsilon'}{2}} {\mathcal F}\Bigl({{\phantom{00000} \theta_1\phantom{hhhhhhh00}\textstyle{\frac 1 2}\phantom{hhhhhhhh000}\theta_t\phantom{hhh}} \atop {\theta_\infty-\textstyle{\frac{\varepsilon}{2}} \phantom{h0}\sigma+n+\textstyle{\frac 1 2} \phantom{h0}\sigma+n+\textstyle{\frac{1-\varepsilon'}{2}} \phantom{h0}\theta_0 }} ;q^{2\theta_t},q^{2\theta_t+2\theta_1}x,t\Bigr)\,, \nonumber\\ &Y^0_{\varepsilon,\varepsilon'}(x,t)= \frac{x^{-\theta_t-\theta_1}}{k_0(\varepsilon)} \sum_{n\in{\mathbb Z}}s^n {\mathcal F}\Bigl( {{\phantom{00000}\theta_1\phantom{hhhhh00} \theta_t\phantom{hhhh00}\textstyle{\frac 1 2}\phantom{hhhh}} \atop {\theta_\infty-\textstyle{\frac{\varepsilon}{2}} \phantom{h0} \sigma+n+\textstyle{\frac 1 2} \phantom{h0} \theta_0+\textstyle{\frac{\varepsilon'}{2}} \phantom{h0} \theta_0}} ;q^{2\theta_t},t,q^{2\theta_t+2\theta_1}x\Bigr) \,, \label{CFT-Y0} \end{aligned}\] where \[\begin{aligned}
&k_\infty(\varepsilon)=q^{(\theta_\infty-\varepsilon/2)^2-2\varepsilon\theta_\infty(\theta_t+\theta_1)} {\mathcal N}'\Bigl({{\phantom{h} \textstyle{\frac 1 2} \phantom{h_3}} \atop{\theta_\infty-\textstyle{\frac{\varepsilon}{2}} \phantom{h_2} \theta_\infty }} \Bigr) \hat{\tau}\,, \\ &k_{0t}(\varepsilon)=k_\infty(\varepsilon) q^{\theta_1^2+\theta_1/2}, \quad k_{0}(\varepsilon)=k_{0t}(\varepsilon) q^{\theta_t^2+\theta_t/2+2\theta_1\theta_t}t^{-\theta_t}, \end{aligned}\]\
and \[\begin{aligned}
&\hat{\tau}= \sum_{n\in{\mathbb Z}}s^n {\mathcal F}\Bigl({{\phantom{h00}\theta_1 \phantom{hhhh} \theta_t\phantom{hhhh}} \atop {\theta_\infty\phantom{h0}\sigma+n\phantom{h0}\theta_0}} ;q^{2\theta_t},t\Bigr) \. \end{aligned}\] We assume that these series converge in the domains [\[domain\]](#domain){reference-type="eqref" reference="domain"}. The asymptotic behavior [\[Yinf\]](#Yinf){reference-type="eqref" reference="Yinf"}, [\[Y0\]](#Y0){reference-type="eqref" reference="Y0"} are simple consequences of the expansion [\[block\]](#block){reference-type="eqref" reference="block"} of conformal blocks. The crucial point is the validity of the connection formulas [\[connection1\]](#connection1){reference-type="eqref" reference="connection1"}, [\[connection2\]](#connection2){reference-type="eqref" reference="connection2"} due to the periodicity property [\[braid-period\]](#braid-period){reference-type="eqref" reference="braid-period"} of the braiding matrix. Therefore, formulas [\[CFT-Yinf\]](#CFT-Yinf){reference-type="eqref" reference="CFT-Yinf"}, [\[CFT-Y0t\]](#CFT-Y0t){reference-type="eqref" reference="CFT-Y0t"}, [\[CFT-Y0\]](#CFT-Y0){reference-type="eqref" reference="CFT-Y0"} solve the generalized Riemann problem. Define the tau function by \[\begin{aligned}
&\tau\left[ \begin{matrix} \theta_1 & \theta_t \\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl|s,\sigma,t\right] =q^{-2(\theta_t+\theta_1)\theta_\infty^2+2\theta_t\theta_1^2} G_q(1+2\theta_\infty)G_q(1-2\theta_0)\cdot \hat{\tau} \. \end{aligned}\] Explicitly it has the AGT series representation \[\begin{aligned}
&\tau\left[ \begin{matrix} \theta_1 & \theta_t \\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl|s,\sigma,t\right] =\sum_{n\in{\mathbb Z}} s^n t^{(\sigma+n)^2-\theta_t^2-\theta_0^2} C\left[\begin{matrix} \theta_1 & \theta_t \\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl|\sigma+n\right] Z\left[\begin{matrix} \theta_1 & \theta_t\\ \theta_\infty &\theta_0 \\ \end{matrix}\Bigl|\sigma+n,t\right]\,, \label{taufn} \end{aligned}\] with the definition \[\begin{aligned}
&C\left[\begin{matrix} \theta_1 & \theta_t \\ \theta_\infty & \theta_0 \\ \end{matrix}\Bigl| \sigma\right] =\frac{\prod_{\varepsilon,\varepsilon'=\pm}G_q(1+\varepsilon\theta_\infty-\theta_1+\varepsilon'\sigma) G_q(1+\varepsilon\sigma-\theta_t+\varepsilon'\theta_0)} {G_q(1+2\sigma)G_q(1-2\sigma)} \,, \\ &Z\left[\begin{matrix} \theta_1 & \theta_t\\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl| \sigma,t\right] =\sum_{\boldsymbol{\lambda}=(\lambda_+,\lambda_-)\in\Lambda^2} t^{|\boldsymbol{\lambda}|}\cdot \frac{ \prod_{\varepsilon,\varepsilon'=\pm} N_{\emptyset,\lambda_{\varepsilon'}}(q^{\varepsilon\theta_\infty-\theta_1-\varepsilon'\sigma}) N_{\lambda_{\varepsilon},\emptyset}(q^{\varepsilon\sigma-\theta_t-\varepsilon'\theta_0})} {\prod_{\varepsilon,\varepsilon'=\pm}N_{\lambda_\varepsilon,\lambda_{\varepsilon'}}(q^{(\varepsilon-\varepsilon')\sigma})} \. \end{aligned}\] Comparing the asymptotic expansion at \(x=0\) and \(x=\infty\), we can express the \((+,-)\) element of the matrix \(A(x,t)\) and \(B(x,t)\) in terms of tau functions.
In, Mano constructed two parametric local solutions of the \(q\)PVI equation near the critical points \(t=0,\infty\). The above formula for \(y\) extends Mano's asymptotic expansion to all orders. We expect that the formula [\[z-tau\]](#z-tau){reference-type="eqref" reference="z-tau"} for \(z\) can be simplified, see [\[yz-final\]](#yz-final){reference-type="eqref" reference="yz-final"} below.
## Bilinear equations
In, Tsuda and Masuda formulated the tau function of the \(q\)PVI equation to be a function on the weight lattice of \(D^{(1)}_5\) with symmetries under the affine Weyl group. Motivated by, we consider the set of eight tau functions defined by the AGT series [\[taufn\]](#taufn){reference-type="eqref" reference="taufn"}: \[\begin{aligned}
{2} \tau_1 &=\tau\left[ \begin{matrix} \theta_1 & \theta_t \\ \theta_\infty+\textstyle{\frac 1 2} & \theta_0\\ \end{matrix}\Bigl|s,\sigma,t\right]\,, & \tau_2 &=\tau\left[ \begin{matrix} \theta_1 & \theta_t \\ \theta_\infty-\textstyle{\frac 1 2} & \theta_0\\ \end{matrix}\Bigl|s,\sigma,t\right]\,, \\ \tau_3 &=\tau\left[ \begin{matrix} \theta_1 & \theta_t \\ \theta_\infty & \theta_0+\textstyle{\frac 1 2}\\ \end{matrix}\Bigl|s,\sigma+\textstyle{\frac 1 2} ,t\right]\,, & \tau_4 &=\tau\left[ \begin{matrix} \theta_1 & \theta_t \\ \theta_\infty & \theta_0-\textstyle{\frac 1 2}\\ \end{matrix}\Bigl|s,\sigma-\textstyle{\frac 1 2} ,t\right]\,, \\ \tau_5 &=\tau\left[ \begin{matrix} \theta_1-\textstyle{\frac 1 2} & \theta_t \\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl|s,\sigma,t\right]\,, & \tau_6 &=\tau\left[ \begin{matrix} \theta_1+\textstyle{\frac 1 2} & \theta_t \\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl|s,\sigma,t\right]\,, \\ \tau_7 &=\tau\left[ \begin{matrix} \theta_1 & \theta_t-\textstyle{\frac 1 2} \\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl|s,\sigma+\textstyle{\frac 1 2} ,t\right]\,, \quad & \tau_8 &=\tau\left[ \begin{matrix} \theta_1 & \theta_t+\textstyle{\frac 1 2} \\ \theta_\infty & \theta_0\\ \end{matrix}\Bigl|s,\sigma-\textstyle{\frac 1 2} ,t\right]\. \\ \end{aligned}\] For convenience, we shift the parameter \(\theta_\infty\) in the previous section to \(\theta_\infty+1/2\). On the basis of a computer run we put forward:
The bilinear equations [\[bilin-1\]](#bilin-1){reference-type="eqref" reference="bilin-1"}--[\[bilin-8\]](#bilin-8){reference-type="eqref" reference="bilin-8"} were originally proved in in the special case when the tau functions are given by Casorati determinants of basic hypergeometric functions. It would be interesting if one can prove them in the general case by a proper extension of the methods of or. | {'timestamp': '2017-08-18T02:07:21', 'yymm': '1706', 'arxiv_id': '1706.01940', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01940'} |
# Introduction {#section:Introduction}
Motivated partially by the findings of , in particular by our difficulties in evolving gravitational wave data close to the critical threshold of black hole formation with the moving puncture gauge, we turned to an alternative formulation and a more accurate numerical method. We implemented the generalized harmonic formulation in a pseudospectral code, `bamps`, which was recently described in detail in . Presently we use this tool to evolve Brill wave initial data in the form most often treated numerically. Primarily we choose such data for ease of comparison with the literature, but additionally since it is axisymmetric it allows us to run the code most efficiently. Ultimately we hope to obtain a proper understanding of, and a robust numerical method for gravitational waves close to the threshold of black hole formation. This study, like , is another step in that direction. The key results in the literature on critical collapse of gravitational waves are those of Abrahams and Evans , who considered one-parameter families of Teukolsky wave initial data in axisymmetry, and found that the resulting black hole mass scales as a power law in a neighborhood of the critical threshold. They also found more tentative evidence for 'echoing', or periodicity in spacetime scale, of the solution. Primarily because of its simplicity, most subsequent studies have focused on Brill wave initial data. In evolutions of Brill waves Sorkin found evidence for *another* critical solution in which the peak of the curvature appears on concentric rings around the symmetry axis. This study employed the generalized harmonic formulation in axisymmetry, and so is the natural starting point for us. For a detailed discussion of critical phenomena in gravitational collapse see . The paper is structured as follows. In section [2](#section:Setup){reference-type="ref" reference="section:Setup"}, we summarize the main features of the `bamps` code and the continuum equations. In section [3](#section:Brill_pos){reference-type="ref" reference="section:Brill_pos"} we present our evolutions. Afterwards we conclude in section [4](#section:Conclusions){reference-type="ref" reference="section:Conclusions"}.
# Setup {#section:Setup}
The `bamps` code uses a pseudospectral method to evolve a first order formulation of the generalized harmonic gauge formulation, with \[\begin{aligned}
\partial_t\alpha&=-\alpha^2 K+\eta_L\alpha^2\log\left(\tfrac{\gamma^{p/2}} {\alpha}\right)+\beta^i\partial_i\alpha\,,\nonumber\\ \partial_t\beta^i&=\alpha^2\,{}^{\textrm{\tiny{(3)}}}\Gamma^i-\alpha\,\partial^i\alpha-\eta_S\beta^i+\beta^j\partial_j\beta^i\,,\label{eqn:3+1_ghg} \end{aligned}\] in the standard \(3+1\) notation. We parametrize the free scalars by \(\eta_L=\bar{\eta}_L\alpha^q\) and \(\eta_S=\bar{\eta}_S\alpha^r\), with \(\bar{\eta}_L,\bar{\eta}_S,p,q\) and \(r=0\) constant. We employ radiation controlling constraint preserving boundary conditions like those described in , imposed via the Bjørhus method , but modified to minimize reflections caused by the use of constraint damping, which can otherwise cause the code to crash with our gauge conditions. The numerical method is similar to that of SpEC , employing many subpatches across which data is communicated using a penalty approach. For our grids we presently take either cubed-sphere or cubed shells . We discretize in space using Chebyschev polynomials, filtering the highest order coefficients . Although `bamps` is fully 3d, to evolve efficiently in axisymmetry we have implemented the analytic-cartoon method , in which all angular spatial derivatives are evaluated using the Killing vector. The resulting regularity conditions are imposed on axis and at the origin. The code is parallelized with MPI at the subpatch level, with one or more subpatches per core, resulting in excellent strong-scaling for up to several thousand cores. Presently we evolve Brill wave initial data , in which the spatial metric takes the form, \[\begin{aligned}
\textrm{d}l^2&=\gamma_{ij}\textrm{d}x^i\textrm{d}x^j =\Psi^4\big[e^{2q}(\textrm{d}\rho^2+\textrm{d}z^2) +\rho^2\textrm{d}\phi^2\big]\,,\label{eqn:Brill_ansatz} \end{aligned}\] and the extrinsic curvature vanishes. We choose always the seed function \(q\) of , \[\begin{aligned}
q(\rho,z)&=A\rho^2e^{-[(\rho-\rho_0)^2/\sigma_\rho^2+(z-z_0)^2/\sigma_z^2]}\,, \label{eqn:brill_seed} \end{aligned}\] with \(A>0\), \(\sigma_\rho=\sigma_z=1\) and \(\rho_0=z_0=0\). We call this data a centered geometrically prolate Brill wave. To build initial data we use the solver of . Our main tool for classifying spacetimes as supercritical is the axisymmetric apparent horizon finder `AHloc`, which we run in postprocessing. A detailed description of the setup can be found in .
# Centered Geometrically prolate Brill waves {#section:Brill_pos}
In this section we present our evolutions of centered geometrically prolate Brill waves. The first numerical evolutions of Brill waves that we are aware of were presented in . Since then Brill wave initial data have been considered multiple times in the numerical relativity literature . Therefore to ensure that `bamps` is performing properly we start with a detailed comparison of the evolutions performed with the seed function [\[eqn:brill_seed\]](#eqn:brill_seed){reference-type="eqref" reference="eqn:brill_seed"} away from criticality. We then compare with the results of before going towards criticality.
## Review and comparison with earlier work {#subsection:comparison}
#### Alcubierre et. al.: {#alcubierre-et.-al. .unnumbered}
In Brill waves were evolved numerically for the first time in 3d. The BSSNOK formulation was used in combination with maximal slicing and vanishing shift. Using the given data [\[eqn:brill_seed\]](#eqn:brill_seed){reference-type="eqref" reference="eqn:brill_seed"} with \(\rho_0=z_0=0\), \(\sigma_\rho=\sigma_\rho=1\), which we consider throughout, it was found that the critical point lies between \(A=4\) and \(A=6\), and furthermore that this finding could be refined to \(A=4.85\pm0.15\), although the data for this latter claim were never presented. Supercriticality was diagnosed by finding an apparent horizon, which occurred for the \(A=6\) data at \(t=7.7\). In it was shown that BSSNOK combined with this gauge choice results in an ill-posed PDE system, meaning that this approach should be either abandoned or modified if we are to achieve accurate results that converge to the continuum solution as resolution is increased.
#### Garfinkle and Duncan: {#garfinkle-and-duncan .unnumbered}
In it was found, evolving Brill wave initial data with \(q\) as in [\[eqn:brill_seed\]](#eqn:brill_seed){reference-type="eqref" reference="eqn:brill_seed"}, again taking \(\rho_0=z_0=0\), \(\sigma_\rho=\sigma_\rho=1\), that the critical amplitude \(A_\star\) lies between \(A=4\) and \(A=6\). The data was classified either by evolving until the spacetime was close to flat and subsequent collapse seemed implausible, or by explicitly finding an apparent horizon. The formulation employed was explicitly axisymmetric, and consisted of a mixed elliptic-hyperbolic system with maximal slicing \(K=0\), well-posedness of which, to the best of our knowledge, has not been studied. We agree with the findings of both and . Because our method employs a different gauge however, it is difficult to make a side-by-side comparison beyond classifying the spacetimes as sub or supercritical. The effect of changing the shape of the initial data parameters \(\sigma_\rho\) and \(\sigma_z\) was also studied in , but we have not yet followed up on this.
#### Rinne: {#rinne .unnumbered}
In the PhD thesis evolutions of centered geometrically prolate Brill waves were presented with a free-evolution and a partially constrained scheme, both in explicit axisymmetry. We focus on the free-evolution scheme, the Z\((2+1)+1\) formulation, since that is where we are able to make the clearest comparisons. In that case harmonic slicing was taken with vanishing shift. This choice is convenient for our comparison because although we can not choose the same shift condition, harmonic lapse is a pure slicing condition, which means that we should obtain the same foliation of the same spacetime (starting from the same initial lapse) albeit with different spatial coordinates. Since there is a preferred observer, namely that at the origin, we can compare quantities explicitly there. Fortunately the work contains several plots along this worldline. In the upper left panel of Fig. [\[fig:Rinne_A1_Kretsch_Lapse_Kretsch\]](#fig:Rinne_A1_Kretsch_Lapse_Kretsch){reference-type="ref" reference="fig:Rinne_A1_Kretsch_Lapse_Kretsch"} we plot the Kretschmann scalar \[\begin{aligned}
I=R_{abcd}R^{abcd}\,, \end{aligned}\] at the origin as a function of time for \(A=1\) centered Brill wave data, which should be compared with Fig. 9.2 of . In this test we evolved with harmonic slicing \(\eta_L=0\) and the damped harmonic shift \(\eta_S=6\) and otherwise our standard setup. The agreement, at least by eye, is extremely good. Taking \(A=4\) it was found that with sufficient resolution a sharp feature in the gradient of the lapse could be resolved. It was found that the data was, in agreement with , subcritical. We see the same result. In the upper right panel of Fig. [\[fig:Rinne_A1_Kretsch_Lapse_Kretsch\]](#fig:Rinne_A1_Kretsch_Lapse_Kretsch){reference-type="ref" reference="fig:Rinne_A1_Kretsch_Lapse_Kretsch"} we show \(\partial_\rho\ln\alpha\) at \(t=1.72\). This is the time at which Rinne finds the largest peak in this quantity. A similar plot is Fig. \(9.3\) in , which can not be directly compared because of the differing spatial coordinates, although the qualitative agreement is very clear. We find that the largest peak appears at around \(t=2.08\), but the magnitudes in \(\partial_\rho\ln\alpha\) differ by less than \(5\%\) across these times. In the lower two panels of Fig. [\[fig:Rinne_A1_Kretsch_Lapse_Kretsch\]](#fig:Rinne_A1_Kretsch_Lapse_Kretsch){reference-type="ref" reference="fig:Rinne_A1_Kretsch_Lapse_Kretsch"}, to be compared with Fig. \(9.9\) of , the left panel shows the logarithm of the lapse at the origin as a function of coordinate time for amplitudes \(A=4,5,6\). In the right hand panel we plot the value of the Kretschmann scalar at the origin. At lower resolutions we find that the Kretschmann scalar exhibits high-frequency oscillatory behavior, but that these wiggles converge away rapidly. The most challenging data evolved in Rinne's experiments was the \(A=5\) wave, for which sub and supercriticality was not discerned using the free-evolution algorithm, partially because the data was still oscillatory at the time the method failed at around \(t=6\). All of the different resolutions we tried with \(A=5\) in this suite of tests crashed at coordinate time \(t\simeq12\). We ran our apparent horizon finder on this data, the result of which is plotted in Fig. [\[fig:Rinne_Horizon\]](#fig:Rinne_Horizon){reference-type="ref" reference="fig:Rinne_Horizon"}. We find the apparent horizon first at \(t\simeq6.2\) and thereafter until the evolution fails. Rinne correctly concluded that the critical amplitude lies below \(A=6\), although no apparent horizon could be found in his data. Instead his classification was made by observing that the Kretschmann scalar was blowing-up. This diagnostic is flawed because as one approaches the critical point we expect to generate arbitrarily large curvature scalars even in subcritical data. In the absence of an apparent horizon or event horizon however, other diagnostics may be similarly flawed, and the Kretschmann scalar is at least a spacetime scalar, so one might prefer it as a diagnostic to 'collapse of the lapse' which is clearly gauge dependent. Running our apparent horizon finder on the \(A=6\) data we find an apparent horizon at \(t\simeq2\) and later.
#### Previous studies with BAM: {#previous-studies-with-bam .unnumbered}
In \(2005\) the BAM finite differencing code was used to evolve centered geometrically prolate Brill waves with the BSSNOK formulation combined with several different gauge conditions, maximal slicing, harmonic slicing, and the moving puncture gauge condition. The main complications were reported to be constraint violation, which was likely caused by lack of resolution, and which did vary significantly from one gauge to another. To understand how much further, if at all, standard modern numerical relativity methods can go beyond those previously discussed, recently in the BAM code was once again used, this time alongside the code of with the focus purely on the moving puncture gauge. Starting with \(A=1\) (that is, weak) data, it was found that the lapse initially decreased, but rapidly returned back to unity; we find qualitatively the same behavior despite the different gauge used in `bamps`, although the lapse function decreases by a smaller amount in the new data. The Kretschmann scalar decreases from a maximum of about 216 at \(t=0\) to zero and reaches a second maximum at \(t=0.7\). The maximum value of the Kretschmann scalar in the domain immediately decreases and never grows beyond the initial value as the wave propagates away. In the upper left panel of Fig. [\[fig:Rinne_A1_Kretsch_Lapse_Kretsch\]](#fig:Rinne_A1_Kretsch_Lapse_Kretsch){reference-type="ref" reference="fig:Rinne_A1_Kretsch_Lapse_Kretsch"} we also plot the central value of the Kretschmann scalar obtained in this BAM experiment. Since the time coordinates used differ, the horizontal axes would be different, but the values of the maxima should be the same. The BAM value is about \(I_0\sim75\) which agrees extremely well with the `bamps` experiment. The Hamiltonian constraint violation in this particular BAM run is of the order \(10^{-3}\) at the time of the second local in time maximum, whereas the roughly analogous \(F_t\) constraint inside `bamps` is less than \(10^{-6}\). This is not a fair comparison, because we are not considering computational cost whatsoever, but does indicate that the `bamps` data is superior in this case. Therefore one expects that as more resolution is added to the BAM grid the result would converge to the `bamps` result. Taking data that is stronger, for example \(A=5\), the method of failed as an incoming pulse in the lapse became evermore sharp, resulting in what seemed to be a coordinate singularity. This would be acceptable if an apparent horizon could be found before the code crashed, but this was not the case. Going to higher amplitudes still, similar failures occurred, and the conclusion was drawn that moving puncture coordinates were not suitable for managing this initial data. We have already seen in our comparison with that using `bamps` the \(A=5\) data can be classified supercritical, and we did not see any sign of a coordinate singularity before an apparent horizon was discovered despite the two lapse functions appearing qualitatively similar at the beginning.
## Comparison with Sorkin
Having collected a bank of evidence that our numerical results are correct we now compare with . We demonstrate firstly that we can obtain qualitatively the same type of behavior described therein, namely that the peak of the Kretschmann scalar appears away from the symmetry axis. This we achieve however just by evolving supercritical Brill wave data and changing the gauge source function we use to evolve it. We find that the position of the peak of the Kretschmann scalar can be controlled by the choice of gauge source function. This can be understood geometrically. Secondly, by locating an apparent horizon in the time development, we demonstrate explicitly that the amplitude \(A=6.073\), evolved and classified in as subcritical, is in fact supercritical. We looked at evolutions of this data at several resolutions, both of the numerical spacetime and the apparent horizon computed on top of the data, and find that the outcome is robust.
#### Simulation setup: {#simulation-setup .unnumbered}
The simulations of this subsection have been carried out on a cubed-ball grid, as described in , with the following setup. The inner cube extends from \(r_{\text{cu}}=0.5\) to \(-r_{\text{cu}}\) and is divided into \(\mathcal{N}_{\text{cu}}=9\) subpatches with \(N_{\text{cu}}=21\) gridpoints in each dimension. For the transition shell from \(r_{\text{cu}}\) to \(r_{\text{cs}}=1.0\) we use only one shell \(\mathcal{N}_{\text{cs}}=1\) with \(N_{\text{cs}}=35\) points in the radial direction. From here we go to the outer boundary at \(r_{\text{ss}}=12\) using \(\mathcal{N}_{\text{ss}}=22\) outer shells with also \(N_{\text{ss}}=35\) radial collocation points. These tests were carried out in 3d using the octant symmetry mode of `bamps`.
#### Position of the peak curvature: {#position-of-the-peak-curvature .unnumbered}
Consider a collapse spacetime with the standard causal structure. Different foliations of the spacetime, in which the time coordinate tends to tick more or less slowly in a region of high curvature depending on some *singularity avoidance* parameter will have different profiles in a spacetime diagram. If we are given a patch of this spacetime up to a finite time coordinate, as in a numerical relativity simulation, the specific observer that encounters the largest curvature before the code crashes depends, among other things, on the singularity avoidance parameter. In our case such a parameter is given by \(\eta_L\). We performed evolutions of a centered Brill wave with \(A=6.073\). This amplitude is shown to be supercritical in the next paragraph. We evolved with fixed \(\eta_S=6,p=1\) and one of \(\eta_L=0,0.2\) or \(\eta_L=0.4\). In the left plot of Fig. [\[fig:Sor_AH\]](#fig:Sor_AH){reference-type="ref" reference="fig:Sor_AH"} we show the lapse in the evolutions at coordinate times \(t=1\), which demonstrate the effect of the singularity avoidance parameter \(\eta_L\). In the pure harmonic slicing case \(\eta_L=0\) we have the strongest singularity avoidance, and find that the peak of the Kretschmann scalar appears at \(\rho=0.88\), where it simply grows until the numerics fail. Increasing the parameter to \(\eta_L=0.2,0.4\) we find that the peak of the Kretschmann appears at a coordinate radius of \(\rho=0.73\) and \(\rho=0.64\) respectively, before the numerics fail.
#### Apparent horizon formation: {#apparent-horizon-formation .unnumbered}
We evolved the same centered \(A=6.073\) data with different resolutions, fixing the gauge parameters \(\eta_S=0.4,p=1\) and \(\eta_S=6\), so the largest choice of \(\eta_L\) from above. We find rapid convergence of the constraints. In particular we used our standard cubed-sphere setup with \(N=21^3\), \(25^3\) and \(29^3\) points per subpatch, and find that, for example, the maximum of the \(C_x\) component of the Harmonic constraints along the x-axis are approximately \(5\times 10^{-5},9\times 10^{-6}\) and \(8\times 10^{-7}\) respectively at \(t\approx1.25\). We then searched for apparent horizons using the method described in . We first find an apparent horizon at around \(t=1\). On a fixed numerical spacetime data set we find perfect fourth order convergence in the apparent horizon data consistent with the Runge-Kutta method employed. Comparing the apparent horizons discovered on the different data we find perfect qualitative agreement. Furthermore we see behavior consistent with rapid convergence when evaluating the differences between the discovered horizons. The apparent horizons of different time slices are plotted in Fig. [\[fig:Sor_AH\]](#fig:Sor_AH){reference-type="ref" reference="fig:Sor_AH"}, where the coordinate expansion of the horizon can clearly be seen. For comparison we again evolved the same initial data inside the BAM finite differencing code, but were unable with our current setup to find apparent horizons from that data. One issue is that the constraint violation is many orders of magnitude greater in the finite differencing code. Qualitatively however we find good agreement in the evolution, at least initially.
#### Summary: {#summary .unnumbered}
We are unable to reproduce the results of , despite using, to the best of our knowledge, identical initial data and gauge. In fact our results appear to contradict the earlier study. The reason for the disagreement is presently not clear. It is possible that an apparent horizon search on Sorkin's older finite differencing data was too challenging because of numerical error, or perhaps even that the numerical dissipation was sufficient to let these strong data spuriously settle down to flat-space, although the latter does not seem likely. The power-law scaling obtained in in the rapid oscillations of the curvature is nevertheless interesting, and would be good to properly understand in the future. Given our earlier code validation and the literature comparison in section [3.1](#subsection:comparison){reference-type="ref" reference="subsection:comparison"} however, we have no reason to doubt the `bamps` results, so for now we move on.
## Discussion {#subsection:Discussion}
#### Critical collapse of the scalar-field: {#critical-collapse-of-the-scalar-field .unnumbered}
Before moving to harder experiments, consider the case of the minimally coupled massless scalar field. Working in spherical symmetry, evolving families of initial data with strength parameter \(A\), Choptuik found compelling numerical evidence for the existence of a critical solution at \(A=A_\star\) serving as the boundary between dispersion and collapse. The appearance of critical phenomena with \(A\) in a neighborhood of \(A_\star\) was neatly explained by the conjecture that the critical solution is an attractor of co-dimension one in phase space. In other words in a neighborhood of the critical solution there should to be just one growing mode. Working in perturbation theory around the critical solution, but crucially allowing for aspherical mode perturbations, Gundlach and Martín-Gárcía found strong numerical evidence for this conjecture . They found that the most slowly decaying mode came with an eigenvalue of \(-\lambda_1\approx 0.02\) associated with a \(Y_{20}\) spherical harmonic. For this matter model the single unstable mode has eigenvalue \(\lambda_0\approx 2.7\). Because of the exponential decay of the former it was conjectured that the qualitative picture obtained in spherical symmetry using the full Einstein equations would not be altered for generic initial data close to criticality. This picture, roughly speaking, says that in a neighborhood of critical collapse the fields should strongly interact in a confined region for a finite, but ever longer time as the critical solution is approached, and ultimately either collapse or disperse. Interestingly for the current study, Choptuik and collaborators then studied axisymmetric configurations and found evidence of a second growing mode associated with a \(Y_{20}\) spherical harmonic, causing the strong field solution to bifurcate into *two* strongly interacting regions, in apparent contradiction with the earlier perturbative result.
#### Issues with the apparent horizon as a diagnostic: {#issues-with-the-apparent-horizon-as-a-diagnostic .unnumbered}
The apparent horizon is used for classification of the spacetime as sub or supercritical. In earlier work the apparent horizon mass has also been used to indicate power-law scaling in the critical regime. Therefore it is worth noting explicitly the weaknesses of this approach. Firstly, the appearance of an apparent horizon guarantees the existence of an event horizon only in strongly asymptotically predictable spacetimes, which arise from generic asymptotically flat initial data only if the weak-cosmic censorship conjecture holds . This fact makes our approach blind to violations of weak-cosmic censorship. It is also possible that no apparent horizon appears in our particular foliation, even if there is a black hole region . Secondly the foliation dependence of the apparent horizon makes it difficult to trust the black hole masses obtained from the horizon area, particularly away from spherical symmetry. We may try to diagnose power-law scaling by looking at horizon mass, but the foliation could be such that the apparent horizon always forms with large area, and we have to choose *when* to evaluate the mass for each member of the one-parameter family. In fact in preliminary experiments we found that when approaching the critical regime, behavior resembling power-law scaling in the initial horizon masses could appear with a particular choice of gauge source function, but then completely disappear once we altered the choice to avoid coordinate singularities. These difficulties make us strongly prefer to study the subcritical approach to black hole formation, because there we can, as in , unambiguously consider the spacetime maximum of the Kretschmann scalar, provided that we are able to evolve the spacetime long enough to be confident that it is really subcritical.
## Towards the critical regime | {'timestamp': '2017-06-07T02:08:10', 'yymm': '1706', 'arxiv_id': '1706.01829', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01829'} |
null | null |
# Introduction
An optimization problem principal to machine learning and statistics is that of finite sums: \[\begin{aligned}
\label{opt:fsm} \min_{\bw\in\reals^d} F(\bw) \coloneqq \frac1n\sum_{i=1}^n f_i(\bw), \end{aligned}\] where the individual functions \(f_i\) are assumed to possess some favorable analytical properties, such as Lipschitz-continuity, smoothness or strong convexity (see for details). We measure the *iteration complexity* of a given optimization algorithm by determining how many evaluations of individual functions (via some external oracle procedure, along with their gradient, Hessian, etc.) are needed in order to obtain an \(\epsilon\)-solution, i.e., a point \(\bw\in\reals^d\) which satisfies \(\bE[F(\bw)-\min_{\bw\in\reals^d} F(\bw)]<\epsilon\) (where the expectation is taken w.r.t. the algorithm and the oracle randomness). Arguably, the simplest way of minimizing finite sum problems is by using optimization algorithms for general optimization problems. For concreteness of the following discussion, let us assume for the moment that the individual functions are \(L\)-smooth and \(\mu\)-strongly convex. In this case, by applying vanilla Gradient Descent (GD) or Accelerated Gradient Descent (AGD, ), one obtains iteration complexity of \[\begin{aligned}
\label{opt:gd_agd} \bigtO{n\kappa\ln(1/\epsilon)} \text{ or } \bigtO{n\sqrt\kappa\ln(1/\epsilon)}, \end{aligned}\] respectively, where \(\kappa\coloneqq L/\mu\) denotes the *condition number* of the problem and \(\tilde{\cO}\) hides logarithmic factors in the problem parameters. However, whereas such bounds enjoy logarithmic dependence on the accuracy level, the multiplicative dependence on \(n\) renders this approach unsuitable for modern applications where \(n\) is very large. A different approach to tackle a finite sum problem is by reformulating it as a stochastic optimization problem, i.e., \(\min_{\bw\in\reals^d} \bE_{i\sim\cU([n])}[f_i(\bw)]\), and then applying a general stochastic method, such as SGD, which allows iteration complexity of \(\bigO{1/\epsilon}\) or \(\bigO{1/\epsilon^2}\) (depending on the problem parameters). These methods offer rates which do not depend on \(n\), and are therefore attractive for situations where one seeks for a solution of relatively low accuracy. An evident drawback of these methods is their broad applicability for stochastic optimization problems, which may conflict with the goal of efficiently exploiting the unique noise structure of finite sums (indeed, in the general stochastic setting, these rates cannot be improved, e.g., ). In recent years, a major breakthrough was made when stochastic methods specialized in finite sums (first SAG and SDCA, and then SAGA, SVRG, SDCA without duality, and others) were shown to obtain iteration complexity of \[\begin{aligned}
\label{rates:optimal} \bigtO{(n+\kappa)\ln(1/\epsilon)}. \end{aligned}\] The ability of these algorithms to enjoy both logarithmic dependence on the accuracy parameter and an additive dependence on \(n\) is widely attributed to the fact that the noise of finite sum problems distributes over a finite set of size \(n\). Perhaps surprisingly, in this paper we show that another key ingredient is crucial, namely, a mean of knowing which individual function is being referred to by the oracle at each iteration. In particular, this shows that variance-reduction mechanisms (see, e.g., ) cannot be applied without explicitly knowing the 'identity' of the individual functions. On the more practical side, this result shows that when data augmentation (e.g., ) is done without an explicit enumeration of the added samples, it is impossible to obtain iteration complexity as stated in ([\[rates:optimal\]](#rates:optimal){reference-type="ref" reference="rates:optimal"}, see for relevant upper bounds). Although variance-reduction mechanisms are essential for obtaining an additive dependence on \(n\) (as shown in ([\[rates:optimal\]](#rates:optimal){reference-type="ref" reference="rates:optimal"})), they do not necessarily yield 'accelerated' rates which depend on the square root of the condition number (as shown in ([\[opt:gd_agd\]](#opt:gd_agd){reference-type="ref" reference="opt:gd_agd"}) for AGD). Recently, generic acceleration schemes were used by and accelerated SDCA to obtain iteration complexity of \[\begin{aligned}
\label{optimal_rate} \bigtO{(n+\sqrt{nk})\ln(1/\epsilon)}. \end{aligned}\] The question of whether this rate is optimal was answered affirmatively by. The first category of lower bounds exploits the degree of freedom offered by a \(d\)-(or an infinite-) dimensional space to show that any first-order and a certain class of second-order methods cannot obtain better rates than ([\[optimal_rate\]](#optimal_rate){reference-type="ref" reference="optimal_rate"}) in the regime where the number of iterations is less than \(\bigO{d/n}\). The second category of lower bounds is based on maintaining the complexity of the functional form of the iterates, thereby establishing bounds for first-order and coordinate-descent algorithms whose step sizes are *oblivious* to the problem parameters (e.g., SAG, SAGA, SVRG, SDCA, SDCA without duality) for *any* number of iterations, regardless of \(d\) and \(n\). In this work, we further extend the theory of oblivious finite sum algorithms, by showing that if a first-order and a coordinate-descent oracle are used, then acceleration is not possible without an explicit knowledge of the strong convexity parameter. This implies that in cases where only poor estimation of the strong convexity is available, faster rates may be obtained through 'adaptive' algorithms (see relevant discussions in ). Next, we show that in the smooth and convex case, oblivious finite sum algorithms which, on average, apply the same update rule at each iteration (e.g., SAG, SDCA, SVRG, SVRG++, and typically, other algorithms with a variance-reduction mechanism as described in ), are bound to iteration complexity of \(\Omega(n+L/\epsilon)\), where \(L\) denotes the smoothness parameter (rather than \(\Omega(n+\sqrt{n L/\epsilon})\)). To show this, we employ a restarting scheme (see ) which explicitly introduces the strong convexity parameter into algorithms that are designed for smooth and convex functions. Finally, we use this scheme to establish a tight dimension-free lower bound for smooth and convex finite sums which holds for oblivious algorithms with a first-order and a coordinate-descent oracle. To summarize, our contributions (in order of appearance) are the following:
- In Section [2](#subsection:no_indices){reference-type="ref" reference="subsection:no_indices"}, we prove that in the setting of stochastic optimization, having finitely supported noise (as in finite sum problems) is not sufficient for obtaining linear convergence rates with a linear dependence on \(n\)-one must also know exactly which individual function is being referred to by the oracle at each iteration. Deriving similar results for various settings, we show that SDCA, accelerated SDCA, SAG, SAGA, SVRG, SVRG++ and other finite sum algorithms must have a proper enumeration of the individual functions in order to obtain their stated convergence rate.
- In Section [3.1](#section:framework){reference-type="ref" reference="section:framework"}, we lay the foundations of the framework of general CLI algorithms (see ), which enables us to formally address oblivious algorithms (e.g., when step sizes are scheduled regardless of the function at hand). In section [3.2](#section:no_acceleration){reference-type="ref" reference="section:no_acceleration"}, we improve upon, by showing that (in this generalized framework) the optimal iteration complexity of oblivious, deterministic or stochastic, finite sum algorithms with *both* first-order and coordinate-descent oracles cannot perform better than \(\Omega(n + \kappa\ln(1/\epsilon))\), unless the strong convexity parameter is provided explicitly. In particular, the richer expressiveness power of this framework allows addressing incremental gradient methods, such as Incremental Gradient Descent and Incremental Aggregated Gradient.
- In Section [3.3](#section:stationary){reference-type="ref" reference="section:stationary"}, we show that, in the \(L\)-smooth and convex case, the optimal complexity bound (in terms of the accuracy parameter) of oblivious algorithms whose update rules are (on average) fixed for any iteration is \(\Omega(n+L/\epsilon)\) (rather then \(\tilde{\cO}(n+\sqrt{nL/\epsilon})\), as obtained, e.g., by accelerated SDCA). To show this, we first invoke a restarting scheme (used by ) to explicitly introduce strong convexity into algorithms for finite sums with smooth and convex individuals, and then apply the result derived in Section [3.2](#section:no_acceleration){reference-type="ref" reference="section:no_acceleration"}.
- In Section [3.4](#section:smooth_lower_bounds){reference-type="ref" reference="section:smooth_lower_bounds"}, we use the reduction introduced in Section [3.3](#section:stationary){reference-type="ref" reference="section:stationary"}, to show that the optimal iteration complexity of minimizing \(L\)-smooth and convex finite sums using oblivious algorithms equipped with a first-order and a coordinate-descent oracle is \(\Omega\circpar{n+\sqrt{{n L }/{\epsilon}}}.\)
# The Importance of Individual Identity {#subsection:no_indices}
In the following, we address the stochastic setting of finite sum problems ([\[opt:fsm\]](#opt:fsm){reference-type="ref" reference="opt:fsm"}) where one is equipped with a *stochastic* oracle which, upon receiving a call, returns some individual function chosen uniformly at random and hides its index. We show that not knowing the identity of the function returned by the oracle (as opposed to an *incremental* oracle which addresses the specific individual functions chosen by the user), significantly harms the optimal attainable performance. To this end, we reduce the statistical problem of estimating the bias of a noisy coin into that of optimizing finite sums. This reduction (presented below) makes an extensive use of elementary definitions and tools from information theory, all of which can be found in. First, given \(n\in\bN\), we define the following finite sum problem \[\begin{aligned}
\label{eq:no_indices} F_\sigma \coloneqq \frac{1}{n}\circpar{\frac{n-\sigma}2 f^+ +\frac{n+\sigma}2 f^-}, \end{aligned}\] where \(n\) is w.l.o.g. assumed to be odd, \(\sigma\in\{-1,1\}\) and \(f^+,f^-\) are some functions (to be defined later). We then define the following discrepancy measure between \(F_1\) and \(F_{-1}\) for different values of \(n\) (see also ), \[\begin{aligned}
\label{definition:delta} \delta(n) = \min_{\bw\in\reals^d}\{F_1(\bw) + F_{-1}(\bw)-F_{1}^*-F_{-1}^*\}, \end{aligned}\] where \(F_\sigma^*\coloneqq\inf_{\bw} F_\sigma(\bw)\). It is easy to verify that no solution can be \(\delta(n)/4\)-optimal for both \(F_1\) and \(F_{-1}\), at the same time. Thus, by running a given optimization algorithm long enough to obtain \(\delta(n)/4\)-solution w.h.p., we can deduce the value of \(\sigma\). Also, note that, one can simplify the computation of \(\delta(n)\) by choosing convex \(f^+,f^-\) such that \(f^+(\bw)=f^-(-\bw)\). Indeed, in this case, we have \(F_1(\bw)=F_{-1}(-\bw)\) (in particular, \(F_1^* = F_{-1}^*\)), and since \(F_1(\bw) + F_{-1}(\bw)-F_{1}^*-F_{-1}^*\) is convex, it must attain its minimum at \(\bw=0\), which yields \[\begin{aligned}
\label{eq:delta_simple} \delta(n)&= 2(F_1(\mathbf{0})-F_{1}^*). \end{aligned}\] Next, we let \(\sigma\in\{-1,1\}\) be drawn uniformly at random, and then use the given optimization algorithm to estimate the bias of a random variable \(X\) which, conditioned on \(\sigma\), takes \(+1\) w.p. \(1/2+\sigma/2n\), and \(-1\) w.p. \(1/2-\sigma/2n\). To implement the stochastic oracle described above, conditioned on \(\sigma\), we draw \(k\) i.i.d. copies of \(X\), denoted by \(X_1,\dots,X_k\), and return \(f^-\), if \(X_i=\sigma\), and \(f^+\), otherwise. Now, if \(k\) is such that \[\begin{aligned}
\bE[F_\sigma(\bw^{(k)})-F^*_\sigma \,|\,\sigma]\le \frac{\delta(n)}{40}, \end{aligned}\] for both \(\sigma=-1\) and \(\sigma= 1\), then by Markov inequality, we have that \[\begin{aligned}
\label{ineq:no_indices_markov} \bP\left(F_\sigma(\bw^{(k)})-F^*_\sigma \ge \delta(n)/4\,\middle| \,\sigma\right)\le 1/10 \end{aligned}\] (note that \(F_\sigma(\bw^{(k)})-F^*_\sigma\) is a non-negative random variable). We may now try to guess the value of \(\sigma\) using the following estimator \[\hat\sigma(\bw^{(k)})=\argmin_{\sigma'\in\{-1,1\}} \{F_{\sigma'}(\bw^{(k)})-F^*_{\sigma'}\},\] whose probability of error, as follows by , is \[\begin{aligned}
\label{eq:sigma_error_prob} \prob{\hat\sigma\neq\sigma}\le 1/10. \end{aligned}\] Lastly, we show that the existence of an estimator for \(\sigma\) with high probability of success implies that \(k=\Omega(n^2)\). To this end, note that the corresponding conditional dependence structure of this probabilistic setting can be modeled as follows: \(\sigma \to X_1,\dots,X_k \to \hat\sigma\). Thus, we have \[\begin{aligned}
\label{sigma_entropy_chain} H(\sigma\,|\,X_1,\dots,X_k) \stackrel{(a)}{\le} H(\sigma\,|\,\hat\sigma) \stackrel{(b)}{\le} H_b(\prob{\hat\sigma\neq\sigma}) \stackrel{(c)}{\le} \frac12, \end{aligned}\] where \(H(\cdot)\) and \(H_b(\cdot)\) denote the Shannon entropy function and the binary entropy function, respectively, \((a)\) follows by the data processing inequality (in terms of entropy), \((b)\) follows by Fano's inequality and \((c)\) follows from [\[eq:sigma_error_prob\]](#eq:sigma_error_prob){reference-type="eqref" reference="eq:sigma_error_prob"}. Applying standard entropy identities, we get \[\begin{aligned}
H(\sigma\,|\,X_1,\dots,X_k) &\stackrel{(d)}{=} H(X_1,\dots,X_k\,|\,\sigma) + H(\sigma)-H(X_1,\dots,X_k) \nonumber\\ &\stackrel{(e)}{=} kH(X_1\,|\,\sigma) + 1-H(X_1,\dots,X_k) \nonumber\\ &\stackrel{(f)}{\ge} kH(X_1\,|\,\sigma) + 1-kH(X_1), \end{aligned}\] where \((d)\) follows from Bayes rule, \((e)\) follows by the fact that \(X_i\), conditioned on \(\sigma\), are i.i.d. and \((f)\) follows from the chain rule and the fact that conditioning reduces entropy. Combining this with and rearranging, we have \[\begin{aligned}
k\ge\frac{1}{2(H(X_1)-H(X_1\,|\,\sigma) )}\ge \frac{1}{2\circpar{1/n}^2}=\frac{n^2}{2}, \end{aligned}\] where the last inequality follows from the fact that \(H(X_1)=1\) and the following estimation for the binary entropy function: \(H_b(p)\ge 1-4\circpar{p-1/2}^2\) (see , Appendix [\[section:technical_lemmas\]](#section:technical_lemmas){reference-type="ref" reference="section:technical_lemmas"}). Thus, we arrive at the following statement.
Instantiating this schemes for \(f^+,f^-\) of various analytical properties yields the following.
A few conclusions can be readily made from . First, if a given optimization algorithm obtains an iteration complexity of an order of \(c(n,\kappa)\ln(1/\epsilon)\), up to logarithmic factors (including the norm of the minimizer which, in our construction, is of an order of \(1/n\) and coupled with the accuracy parameter), for solving smooth and strongly convex finite sum problems with a stochastic oracle, then \[\begin{aligned}
c(n,\kappa)&= {\tilde\Omega\circpar{\frac{n^2}{\ln(n^2/(\kappa+1))}}}. \end{aligned}\] Thus, the following holds for optimization of finite sums with smooth and strongly convex individuals.
This implies that variance-reduction methods such as, SAG, SAGA, SDCA and SVRG (possibly combining with acceleration schemes), which exhibit linear dependence on \(n\), cannot be applied when data augmentation is used. In general, this conclusion also holds for cases when one applies general first-order optimization algorithms, such as AGD, on finite sums, as this typically results in a linear dependence on \(n\). Secondly, if a given optimization algorithm obtains an iteration complexity of an order of \(n + {L^\beta\|\bw^{(0)}-\bw^*\|^2}/{\epsilon^\alpha}\) for solving smooth and convex finite sum problems with a stochastic oracle, then \(n+L^{\beta-\alpha} n^{2(\alpha-1)} = \Omega(n^2)\). Therefore, \(\beta=\alpha\) and \(\alpha\ge2\), indicating that an iteration complexity of an order of \(n+L\|\bw^{(0)}-\bw^*\|^2/\epsilon\), as obtained by, e.g., SVRG++, is not attainable with a stochastic oracle. Similar reasoning based on the Lipschitz and strongly convex case in shows that the iteration complexity guaranteed by accelerated SDCA is also not attainable in this setting.
# Oblivious Optimization Algorithms {#section:oblivious}
In the previous section, we discussed different situations under which variance-reduction schemes are not applicable. Now, we turn to study under what conditions can one apply acceleration schemes. First, we define the framework of oblivious CLI algorithms. Next, we show that, for this family of algorithms, knowing the strong convexity parameter is crucial for obtaining accelerated rates. We then describe a restarting scheme through which we establish that *stationary* algorithms (whose update rule are, on average, the same for every iteration) for smooth and convex functions are sub-optimal. Finally, we use this reduction to derive a tight lower bound for smooth and convex finite sums on the iteration complexity of any oblivious algorithm (not just stationary).
## Framework {#section:framework}
In the sequel, following, we present the analytic framework through which we derive iteration complexity bounds. This, perhaps pedantic, formulation will allows us to study somewhat subtle distinctions between optimization algorithms. First, we give a rigorous definition for a *class of optimization problems* which emphasizes the role of prior knowledge in optimization.
In finite sum problems, \(\cF\) comprises of functions as defined in ([\[opt:fsm\]](#opt:fsm){reference-type="ref" reference="opt:fsm"}); the side-information may contain the smoothness parameter \(L\), the strong convexity parameter \(\mu\) and the number of individual functions \(n\); and the oracle may allow one to query about a specific individual function (as in the case of incremental oracle, and as opposed to the stochastic oracle discussed in Section [2](#subsection:no_indices){reference-type="ref" reference="subsection:no_indices"}). We now turn to define CLI optimization algorithms (see for a more comprehensive discussion).
Throughout the rest of the paper, we shall be interested in oblivious CLI algorithms (for brevity, we usually omit the 'CLI' qualifier) equipped with the following two incremental oracles: \[\begin{aligned}
\label{oracles} &\text{Generalized first-order oracle: } \quad O(\bw;A,B,\bc,i) \coloneqq A\nabla f_i(\bw) + B\bw+\bc, \nonumber \\ &\text{Steepest coordinate-descent oracle:} \quad O(\bw;j,i) \coloneqq \bw + t^*\be_j, \end{aligned}\] where \(A,B\in\reals^{d\times d},\bc\in\reals^d, i\in[n],j\in[d]\), \(\be_j\) denotes the \(j\)'th \(d\)-dimensional unit vector and \(t^*\in \argmin_{t\in\reals} f_j(w_1,\dots,w_{j-1},w_j+t,w_{j+1},\dots,w_d)\). We restrict the oracle parameters such that only one individual function is allowed to be accessed at each iteration. We remark that the family of oblivious algorithms with a first-order and a coordinate-descent oracle is wide and subsumes SAG, SAGA, SDCA, SDCA without duality, SVRG, SVRG++ to name a few. Also, note that coordinate-descent steps w.r.t. partial gradients can be implemented using the generalized first-order oracle by setting \(A\) to be some principal minor of the unit matrix (see, e.g., RDCM in ). Further, similarly to, we allow both first-order and coordinate-descent oracles to be used during the same optimization process.
## No Strong Convexity Parameter, No Acceleration for Finite Sum Problems {#section:no_acceleration}
Having described our analytic approach, we now turn to present some concrete applications. Below, we show that in the absence of a good estimation for the strong convexity parameter, the optimal iteration complexity of oblivious algorithms is \(\Omega(n+k\ln(1/\epsilon))\). Our proof is based on the technique used in (see for a brief introduction of the technique). Given \(0<\epsilon<L\), we define the following set of optimization problems (over \(\reals^d\) with \(d>1\)) \[\begin{aligned}
\label{problem:finite_sum_no_acceleration} F_{\mu}(\bw)&\coloneqq \frac{1}{n}\sum_{i=1}^n \circpar{\frac{1}{2}\bw^\top Q_{\mu} \bw-\bq^\top\bw},\text{where}\\ Q_{\mu}&\coloneqq\mymat{\frac{L+\mu}{2} & \frac{\mu-L}{2}\vspace{0.05cm}\\ \frac{\mu-L}{2}&\frac{L+\mu}{2} \\ && \mu \\ &&& \ddots \\ &&&&\mu }, ~\bq\coloneqq \frac{\epsilon R}{ \sqrt2} \mymat{1\\1\\0\\\vdots\\0},\nonumber \end{aligned}\] parametrized by \(\mu\in(\epsilon,L)\) (note that the individual functions are identical. We elaborate more on this below). It can be easily verified that the condition number of \(F_\mu\), which we denote by \(\kappa(F_\mu)\), is \(L/\mu\), and that the corresponding minimizers are \(\bw^*(\mu) =({\epsilon R}/{\mu\sqrt{2}},{\epsilon R}/{\mu\sqrt{2}}, 0,\dots,0)^\top\) with norm \(\le R\). If we are allowed to use different optimization algorithm for different \(\mu\) in this setting, then we know that the optimal iteration complexity is of an order of \((n+\sqrt{n\kappa(F_\mu)})\ln(1/\epsilon)\). However, if we allowed to use only one single algorithm, then we show that the optimal iteration complexity is of an order of \(n+\kappa(F_\mu)\ln(1/\epsilon)\). The proof goes as follows. First, note that in this setting, the oracles defined in ([\[oracles\]](#oracles){reference-type="ref" reference="oracles"}) take the following form, \[\begin{aligned}
\label{eq:oracles_finite_sum} &\text{Generalized first-order oracle:} ~O(\bw;A,B,\bc,i) = A(Q_{\mu} \bw-\bq) + B\bw+\bc,\\ &\text{Steepest coordinate-descent oracle:} ~O(\bw;j,i) = \circpar{I-(1/(Q_{\mu})_{jj})\be_i(Q_{\mu})_{j,*} }\bw-q_j/(Q_{\mu})_{jj}\be_j.\nonumber \end{aligned}\] Now, since the oracle answers are linear in \(\mu\) and the \(k\)'th iterate is a \(k\)-fold composition of sums of the oracle answers, it follows that \(\bw_1^{(k)}\) forms a \(d\)-dimensional vector of univariate polynomials in \(\mu\) of degree \(\le k\) with (possibly random) coefficients (formally, see , Appendix [\[section:technical_lemmas\]](#section:technical_lemmas){reference-type="ref" reference="section:technical_lemmas"}). Denoting the polynomial of the first coordinate of \(\bE\bw_1^{(k)}(\mu)\) by \(s(\mu)\), we see that for any \(\mu\in(\epsilon,L)\), \[\begin{aligned}
\bE\|\bw_1^{(k)}(\mu)-\bw^*(\mu)\| \ge \|\bE\bw_1^{(k)}(\mu)-\bw^*(\mu)\|\nonumber \ge \left|s(\mu)-\frac{R\epsilon}{\sqrt2 \mu} \right|\nonumber \ge \frac{R\epsilon}{\sqrt{2}L} \left|\frac{\sqrt{2}s(\mu)\mu}{R \epsilon}-1\right|,\nonumber \end{aligned}\] where the first inequality follows by Jensen inequality and the second inequality by focusing on the first coordinate of \(\bE\bw^{(k)}(\eta)\) and \(\bw^*(\eta)\). Lastly, since the coefficients of \(s(\mu )\) do not depend on \(\mu\), we have by in Appendix [\[section:technical_lemmas\]](#section:technical_lemmas){reference-type="ref" reference="section:technical_lemmas"}, that there exists \(\delta>0\), such that for any \(\mu\in(L-\delta,L)\) it holds that \[\begin{aligned}
\frac{R\epsilon}{\sqrt{2}L} \left|\frac{\sqrt{2}s(\mu)\mu}{R \epsilon}-1\right|\ge \frac{R\epsilon}{\sqrt{2}L} \circpar{1-\frac{1}{\kappa(F_\mu)}}^{k+1}, \end{aligned}\] by which we derive the following.
The \(n\) part of the lower bound holds for any type of finite sum algorithm and is proved in. The lower bound stated in is tight up to logarithmic factors and is attained by, e.g., SAG. Although relying on a finite sum with identical individual functions may seem somewhat disappointing, it suggests that some variance-reduction schemes can only give optimal dependence in terms of \(n\), and that obtaining optimal dependence in terms of the condition number need to be done through other (acceleration) mechanisms (e.g., ). Lastly, note that, this bound holds for any number of iterations (regardless of the problem parameters).
## Stationary Algorithms for Smooth and Convex Finite Sums are Sub-optimal {#section:stationary}
In the previous section, we showed that not knowing the strong convexity parameter reduces the optimal attainable iteration complexity. In this section, we use this result to show that whereas general optimization algorithms for smooth and convex finite sum problems obtain iteration complexity of \(\tilde{\cO}(n+\sqrt{nL/\epsilon})\), the optimal iteration complexity of stationary algorithms (whose expected update rules are fixed) is \(\Omega(n+L/\epsilon)\). The proof (presented below) is based on a general restarting scheme (see Scheme 1) used in. The scheme allows one to apply algorithms which are designed for \(L\)-smooth and convex problems on smooth and strongly convex finite sums by explicitly incorporating the strong convexity parameter. The key feature of this reduction is its ability to 'preserve' the exponent of the iteration complexity from an order of \(C(f)(L/\epsilon)^\alpha\) in the non-strongly convex case to an order of \(\circpar{C(f)\kappa}^{\alpha}\ln(1/\epsilon)\) in the strongly convex case, where \(C(f)\) denotes some quantity which may depend on \(f\) but not on \(k\), and \(\alpha\) is some positive constant.
We note that, this lower bound is tight and is attained by, e.g., SDCA.
## A Tight Lower Bound for Smooth and Convex Finite Sums {#section:smooth_lower_bounds}
We now turn to derive a lower bound for finite sum problems with smooth and convex individual functions using the restarting scheme shown in the previous section. Note that, here we allow any oblivious optimization algorithm, not just stationary. The technique shown in Section [3.2](#section:no_acceleration){reference-type="ref" reference="section:no_acceleration"} of reducing an optimization problem into a polynomial approximation problem was used in to derive lower bounds for various settings. The smooth and convex case was proved only for \(n=1\), and a generalization for \(n>1\) seems to reduce to a non-trivial approximation problem. Here, using Scheme 1, we are able to avoid this difficulty by reducing the non-strongly case to the strongly convex case, for which a lower bound for a general \(n\) is known. The proof follows the same lines of the proof of . Given an oblivious optimization algorithm for finite sums with smooth and convex individuals equipped with oracles ([\[oracles\]](#oracles){reference-type="ref" reference="oracles"}), we apply again Scheme 1 to get an algorithm for the smooth and strongly convex case, whose iteration complexity is as in ([\[new_finite_sum_convergence\]](#new_finite_sum_convergence){reference-type="ref" reference="new_finite_sum_convergence"}). Now, crucially, oblivious algorithm are invariant under Scheme 1 (that is, when applied on a given oblivious algorithm, Scheme 1 produces another oblivious algorithm). Therefore, using, we obtain the following.
This bound is tight and is obtained by, e.g., accelerated SDCA. Optimality in terms of \(L\) and \(\epsilon\) can be obtained simply by applying Accelerate Gradient Descent, or alternatively, by using an accelerated version of SVRG as presented in. More generally, one can apply acceleration schemes, e.g.,, to get an optimal dependence on \(\epsilon\). | {'timestamp': '2017-12-08T02:00:54', 'yymm': '1706', 'arxiv_id': '1706.01686', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01686'} |
null | null |
null | null |
null | null |
null | null |
# Introduction
most noticeable developments foreseen in the near future in power systems involve Distribution Networks (DNs). Future DNs are expected to host a big percentage of Renewable Energy Sources (RES) and other Distributed Energy Resources (electric vehicles, flexible loads, fuel cells, batteries, etc.). Moreover, it is expected that DNs will be called upon to actively support the bulk Transmission Network (TN) participating in ancillary services with the help of Information and Communication Technologies (ICT) and advanced management and control techniques. For these reasons, the interaction between the Transmission and Distribution (T&D) grids has become the focus in many research areas of power systems over the last years. It has drawn significant attention in the areas of voltage stability and support , combined system dynamic stability , optimisation of power and reserves , dynamic simulations for security assessment, and much more. This increased research interest has been coupled with an equally high number of national and international research projects, funded both by industry and governmental agencies, to define or analyse the interactions between Transmission System Operators (TSOs) and Distribution System Operators (DSOs), e.g. .
## Review of popular test system models
Several test systems exist for separately studying TNs or DNs. For TN studies, the most widely used systems are the ones developed by IEEE, for power flow and transient stability studies . More specialised test system have also been proposed, for instance, the RTS-79 and RTS-96 for reliability studies, or the revised Nordic system for dynamic and voltage stability studies . Several variants of these models have been published over the years depending on the type of study and the phenomena tackled by the method being tested. Other research groups and organisations have also developed their own test systems; for example, the 150-bus synthetic system from the University of Illinois, Urbana-Champaign , the 9-bus and 179-bus systems from the Western System Coordinating Council , used for transient stability studies , and more . Regarding distribution network models, a wide variety of distribution test systems have been proposed by IEEE and can be found in. In addition, the 33-bus and the 69-bus test systems have been used, among other things, to study the impact of distributed generation while the RBTS 6-bus test system has been used for reliability analysis . Several other, isolated or collective, efforts have been made to prepare and make available generic DN test systems, e.g. . While there is an abundance of individual TN and DN systems, there is a lack of combined T&D systems allowing to investigate the interplay between them. Such systems are required to examine the impact of active distribution networks on the TN, study various proposed methods for DNs to support the TN in steady-state or during fault dynamics, develop or validate DN equivalent models that can later be used in bulk TN stability studies, test T&D co-optimization algorithms (for planning or operation), analyze the techniques for including DNs to power markets, and many more. In the past, specific test systems, such as in , have been developed by individual researchers or groups to analyse the T&D interactions. However, constructing such a system is a tedious task with many challenges and parameters to be selected. In addition, the data of these systems is rarely published and the customization performed to match the specific problem studied, makes them difficult to be used in other applications.
## Paper Structure
The remaining paper is organised as follows. In , the \"template\" TN and DN models that are used to generate the combined T&D system are presented. Then, in , the key parameters for selecting the characteristics of the generated system are explained and in , the procedure of generating the combined T&D system is detailed. Finally, some example test cases are given in followed by some concluding remarks in .
# Template test systems {#sec:systems}
In this section, we briefly present the TN and DN models that are combined to generate the resulting T&D system. The TN is based on the model documented in and sketched in Fig. [\[fig:one_line_tn\]](#fig:one_line_tn){reference-type="ref" reference="fig:one_line_tn"}. It is a variant of the well-known Nordic test system that has been recently revised by the IEEE Power System Dynamic Performance Committee. The system is separated into four zones: *Equiv*, *North*, *Central*, and *South*. Most of the generation is situated in the North and consists of hydropower plants. The rest of the generation, in the *Central* and *South* areas, consists of thermal power plants. Most of the consumption is located in the *Central* and *South* areas. *Equiv* is an equivalent of an external system connected to the North area. In this model, the distribution networks are represented as aggregated loads (see Fig. [\[fig:split\]](#fig:split){reference-type="ref" reference="fig:split"}, left-hand side). The TN model includes 74 buses: 32 at the transmission level, 20 are generator terminal buses, and 22 buses at the distribution level (medium voltage side of distribution transformers) where the aggregated loads are connected. It includes 102 branches, among which 22 are distribution and 20 step-up transformers. The model used for the DNs was developed by the Centre for Sustainable Electricity and Distributed Generation (SEDG) . It represents a radial 11kV urban network fed from a 33kV supply point and its one-line diagram is shown in Fig. [\[fig:one_line_dn\]](#fig:one_line_dn){reference-type="ref" reference="fig:one_line_dn"}. The system has been modified to include two types of DGs and to accommodate higher loading levels. The feeders 1-4 are serving bigger consumers and the DGs connected there are considered to be controllable micro-turbines or small synchronous machines. The remaining feeders are serving residential consumers and the DGs consist of aggregate models of rooftop photovoltaic (PV) systems, thus uncontrolled and generating at their maximum power with unity power factor. Overall, the DN system includes 75 buses with 4 DGs on feeders 1-4 and 18 aggregate PV systems on the remaining feeders.
# Key parameters {#sec:key_parameters}
In this section, we present some key parameters that are used to generate the final test system. These user-defined parameters are located in the file *parameters.m*, at the root of the toolbox.
- *Penetration level*: This value defines the percentage of the active power demand to be generated by the DGs in each DN. It is defined as the ratio between the total active power injected by the DGs and the total power demand of the loads served by the DN, i.e. \[PL = \frac{\sum_{i=1}^{N_{DG}} P_{DGi}}{\sum_{j=1}^{N_{Dj}} P_{D}}\] where \(N_{DG}\) (resp. \(N_{D}\)) is the number of DGs (resp. loads) within the DN. It has to be noted that the penetration level is defined *per DN* and not on the entire combined T&D system. Due to the existence of loads connected at the TN level, the penetration of the combined system will be less than this value.
- *Generation Split*: This parameter defines the way of splitting the generation between the larger DGs (feeders 1-4 in Fig. [\[fig:one_line_dn\]](#fig:one_line_dn){reference-type="ref" reference="fig:one_line_dn"}) and the smaller ones (feeders 5-8 in Fig. [\[fig:one_line_dn\]](#fig:one_line_dn){reference-type="ref" reference="fig:one_line_dn"}). It allows assigning higher generation levels to the controllable units (feeders 1-4) or to the uncontrollable units (feeders 5-8).
- *Constant load*: This parameter defines which possible vision of the future power grid is considered for the generated T&D system. The options are:
- Scenario 1 (*constant load='false'*) assumes that the growth of distributed generation will outpace the growth of electricity demand in DNs. If we consider this from the TN point of view, it means that the net demand of the DNs will decrease. If the penetration level is very high, a reverse power flow can be observed from the DNs to the TN. This option modifies the TN power flows as given in .
- Scenario 2 (*constant load='true'*) assumes that the electricity demand will also increase (e.g., because of electric vehicles) alongside distributed generation. Thus, the excess demand will be covered by the DGs and the TN will see the same total demand from the DNs (hence 'constant load'). This option does not modify the TN power flows as defined in .
- *Random*: Since the same DN template is used to replace all of the TN aggregate loads, this can lead to artificial synchronisations between the various DNs. This setting allows for the parameters of the DNs (e.g., penetration level, generator split, etc.) to be slightly varied around the original values (maximum \(\pm5\%\)), thus introducing some diversity between DNs.
- *Large system*: In the default setting (*Large system='true'*), all the loads of the original TN except the ones in the *Equiv* area (on buses 71 and 72 in Fig. [\[fig:one_line_tn\]](#fig:one_line_tn){reference-type="ref" reference="fig:one_line_tn"}) are systematically replaced by detailed DNs. This leads to a T&D system of approximately 22000 buses. Alternatively, if the parameter is set to *'false'*, then only the loads in the *Central* area are replaced, thus leading to a T&D system of approximately 15000 buses.
- *Oversize*: This option allows to oversize the power consumed by each DN, thus creating several security violations. This is useful when the user needs a severely congested test system to test management and operation techniques relying heavily on DGs. As a consequence of the overloading, the number of detailed DNs required to replace each TN load is decreased and consequently also the number of buses in the combined system.
- *Run OPF*: Determines if the user wants to optimise the operating point of the generated T&D system before exporting the data. The optimisation procedure is further detailed in .
- *Export format*: Defines the format in which the power flow data is exported (MATPOWER or a custom exporter). Two example custom exporters (one for power flow and one for time-domain dynamic simulations) are provided in .
# Methodology {#sec:Methodology}
In this section, the general methodology of generating the combined T&D system is presented. As mentioned earlier, *TDNetGen* systematically replaces the aggregated loads of the original TN system with detailed DNs, based on the template DN model presented in . The procedure is summarised here and detailed below:
1. Initialize the TN model. Compute the amount of load, per TN bus, that needs to be replaced by detailed DNs based on the *Large system* parameter.
2. Initialize the template DN system and find the maximum DN capacity while avoiding voltage violations. Using the maximum DN capacity and the load per TN bus, calculate the minimum number of DNs required to replace the aggregated loads. Generate the corresponding DN models and compute their individual capacity.
3. Using the parameters *Penetration level*, *Generation split*, *Constant load*, *Random*, and *Oversize*, and their capacity customize each DN operating point by scaling the load and the DGs.
4. Generate the combined T&D system by interconnecting the models, treating the naming conventions, and ensuring continuity over the boundaries (distribution transformers).
5. **(Optional)** If the *Run OPF* parameter is selected, optimise the combined T&D system operating point to reduce generation cost and alleviate voltage problems.
6. Export the combined system data in MATPOWER format or one of the custom exporters, as selected by *Export format*.
## Transmission network initialization
To decrease the complexity of the T&D model generation, a master-slave approach is followed, illustrated in Fig. [\[fig:master_slave\]](#fig:master_slave){reference-type="ref" reference="fig:master_slave"}. This approach is only used for the construction of the T&D model. Once the model is exported, the user can analyse it using standard integrated approaches (e.g., MATPOWER's standard power-flow solvers). First, the TN model (located under *input_data/tn_template.mat*) is loaded in MATPOWER and a power-flow solution is performed to compute the voltages at the TN load buses. Then, depending on the *Large system* parameter, it is selected which TN loads will be replaced by detailed DNs. That is, either all the loads in the system will be replaced or only the ones in the Central area. This approach allows the user to modify the TN operating point before calling *TDNetGen* in order to test different operating conditions. For instance, the parameters of the large TN generators could be modified and the TN load consumption changed to create different scenarios. However, the TN load buses and names should remain unchanged as the following procedure depends on these.
## Distribution network initialization
In a second step, the DN template model (located in *input_data/dn_template.mat*) is loaded into MATPOWER. The maximum capacity of the DN is calculated by gradually scaling the DN loads under constant power factor while keeping the output of the DGs to zero and making sure that the voltage constraints are not violated. Finally, if the parameter *oversize* is larger than \(1.0\), all the DN loads are scaled by this factor, ignoring the voltage violations. Then, the number of DNs needed to replace each aggregated TN load (as shown in Fig. [\[fig:split\]](#fig:split){reference-type="ref" reference="fig:split"}) is computed by dividing the TN load consumed power with the DN maximum capacity and rounding up to the nearest integer. The implicit assumptions made in this step are that i) the original TN loads are pure loads without any aggregated distributed generation, and ii) the DNs used to replace the TN loads are operating close to their maximum capacity (or oversized). Finally, using the number of DNs per TN bus, the individual DN models are generated and their load consumption is calculated. Similarly to the TN model, this approach allows the user to modify the DN parameters before executing *TDNetGen*. However, much of the scaling and optimisation procedure that follows relies on the exact naming conventions and topology of the DN (location of DGs), thus these should be kept for the modified system.
## Customization of distribution networks {#sec:DN_custom}
In this step, the created DN models are customised according to the parameters defined by the user. The amount of renewable generation is defined by the *penetration level* and the DN consumption; its allocation is determined by the generation splitting parameter, which allows the user to assign more or less power to the controllable DGs in feeders 1-4. Then, if the *Random* flag has been set by the user, a randomization of the DN parameters is performed to make sure that the DN models are not identical. In addition, if the *constant load* parameter is set to true, the active power demand of the DN models is increased to match the introduced DGs, so that the total demand seen by the TN remains the same before and after the replacement of the TN loads by the DN models. At the end of this step, a set of DN models with the user defined characteristics is available to be integrated into the TN.
## Combined system generation {#sec:combined}
As mentioned previously, the integration of the DNs is achieved by replacing the aggregated TN loads by several DNs in parallel (see Fig. [\[fig:split\]](#fig:split){reference-type="ref" reference="fig:split"}). Before connecting the DN models to the TN, the conditions over the boundary elements (distribution transformers) need to be consistent. That is, the magnitude and angle of the high voltage bus to which the DN will be attached (shown in Fig. [\[fig:one_line_dn\]](#fig:one_line_dn){reference-type="ref" reference="fig:one_line_dn"}) needs to match the voltage used for computing the DN power flows in . Consequently, the angles of the other DN buses need to be adjusted accordingly. Moreover, the distribution transformers between the TN and the DNs (see Fig. [\[fig:one_line_dn\]](#fig:one_line_dn){reference-type="ref" reference="fig:one_line_dn"}) are considered to be equipped with On-Load Tap-Changing (OLTC) controllers and used to regulate the voltage of the DNs. Basically, the transformer ratio is modified in discrete steps to maintain the voltage at the low-voltage side of the transformer within some deadband as follows for the \(i\)-th transformer: \[\label{eq:OLTC} \begin{cases} \mbox{if } V^C_i > V^{set}_i+\frac{DB_i}{2} & r_i=r_i+1 \\ \mbox{if } V^C_i < V^{set}_i-\frac{DB_i}{2} & r_i=r_i-1 \end{cases}\] where \(V^{set}_i\) and \(DB_i\) are the controller setpoint and deadband, respectively; \(V^C_i\) is the controlled voltage; and, \(r_i\) is the discrete tap defining the transformer ratio. After the customization of the DNs in , the DN voltage controlled by the OLTC (\(V^C_i\)) might be out of the deadband and require adjusting the transformer ratio. Unfortunately, MATPOWER does not support OLTC transformers. Thus, their functionality was implemented externally with an iterative procedure: a combined T&D power flow solution is performed, then the voltages of the DN substations are checked and the OLTC ratios adjusted accordingly. This is followed by another combined system power flow solution to update the values. This sequence is repeated until the DN voltage set-points (which can be changed by the user in *parameters.m*) are reached. Some safeguards in the form of a maximum allowed changes have been implemented to avoid infinite cycling of the OLTCs. Finally, the combined T&D network model is generated. In the case of *constant load* set to true, the TN power flows and voltages will be unchanged. Alternatively, the TN power flows will be different and at high penetration levels even reversed. Moreover, due to the integration of the DGs, there might be some voltage violations both in the DNs as well as in the TN. The latter only if *constant load* is set to false.
## Operating point optimization {#sec:OPF}
The resulting T&D system from the previously described procedure might lead to increased generation costs as well as voltage violations. *TDNetGen* provides the option to run an AC OPF to optimise the combined system operation and alleviate any voltage problems before exporting the data. Unlike the previous steps, this is an optional step (controlled by the *Run OPF* parameter) as the users might want to implement their own OPF algorithm or test some operational schemes for alleviating voltage problems. To optimise the operating point, the following OPF problem needs to be solved:
where \(\bm x\) is the vector of state variables (i.e., voltage magnitudes and phases at all buses), \(\bm c\) is an \(n_c\) dimensional vector of continuous control variables (i.e., active and reactive powers of generators) and \(\underline{\bm c}\) (resp. \(\overline{\bm c}\)) is its corresponding vector of lower (resp. upper) bounds, \(r\) is an \(n_d\) dimensional vector of discrete control variables (i.e., the OLTC transformer ratios), \(r_i^j\) is the \(j\)-th discrete value of discrete variable \(r_i\), \(r_i^{max}\) is the number of discrete positions of the OLTC, \(\bm f(\cdot)\) is the objective function, \(\bm g(\cdot)\) and \(\bm h(\cdot)\) are vectors of functions which model equality and inequality constraints. The objective function [\[eq:OPF-obj\]](#eq:OPF-obj){reference-type="eqref" reference="eq:OPF-obj"} is to minimise the cost of supplying the load. Only the large TN generators and the DGs located in feeders 1-4 of each DN are dispatched. That is, the PV systems in feeders 5-8 do not participate and are assumed to always operate at maximum power and unity power factor. The generator quadratic cost functions were taken from and, for the TN and DN generators respectively. Their values are defined in *functions/add_gen_costs_for_OPF.m* and can be modified by the user. The equality constraints [\[eq:OPF-Eq\]](#eq:OPF-Eq){reference-type="eqref" reference="eq:OPF-Eq"} are the AC bus power flow equations, the inequality constraints [\[eq:OPF-Ineq\]](#eq:OPF-Ineq){reference-type="eqref" reference="eq:OPF-Ineq"} refer to operational limits (i.e., the voltage magnitudes security constraints), the inequality constraints [\[eq:OPF-Ineq2\]](#eq:OPF-Ineq2){reference-type="eqref" reference="eq:OPF-Ineq2"} refer to physical limits of equipment (i.e., bounds on generators active/reactive powers), and the constraints [\[eq:OPF-disc\]](#eq:OPF-disc){reference-type="eqref" reference="eq:OPF-disc"} describe the discrete variable values of the OLTC controllers. Efficiently solving a large-scale, mixed-integer, optimisation problem as described by [\[eq:OPF-opti\]](#eq:OPF-opti){reference-type="eqref" reference="eq:OPF-opti"} can be challenging. Several methods have been proposed in the literature based on the round-off strategy, using penalty terms, or several other heuristic methods (several examples can be found in and their references). However, the focus of this paper is on providing a parametrizable, large-scale, test-system and not on developing solution techniques for OPF problems with discrete decision variables. Thus, a simple, heuristic, iterative relaxation method is employed in this work. The procedure used in *TDNetGen* is summarised in Fig. [\[fig:OPF-SOl\]](#fig:OPF-SOl){reference-type="ref" reference="fig:OPF-SOl"}. First, the voltage constraints [\[eq:OPF-Ineq\]](#eq:OPF-Ineq){reference-type="eqref" reference="eq:OPF-Ineq"} are relaxed and the discrete variables [\[eq:OPF-disc\]](#eq:OPF-disc){reference-type="eqref" reference="eq:OPF-disc"} are fixed constant. The latter transforms the mixed-integer OPF into a \"standard AC OPF problem\" that is solved with the included MATPOWER algorithm to compute an estimate of the continuous control variables \(\bm c\). Based on this estimate, the discrete variables \(\bm r\) are updated using the control rules of [\[eq:OLTC\]](#eq:OLTC){reference-type="eqref" reference="eq:OLTC"}. Finally, the security constraints are tightened and the procedure is repeated until the problem is solved with the final constraints, selected by the user, and the OLTC control rules [\[eq:OLTC\]](#eq:OLTC){reference-type="eqref" reference="eq:OLTC"} are satisfied. Starting from MATPOWER version 6.0, a new tool called MOST is also provided for optimal scheduling including uncertainty in demand and RES generation. However, this has not been tested in this work.
## Data exporters {#sec:export}
The final step is to export the combined system into the *output_data* folder. In the default format, *TDNetGen* exports a MATPOWER case file that can be loaded, modified, and analysed, without the need of the *TDNetGen* being present. In addition, the toolbox allows the user to build a custom exporter to any format, depending on the simulation software used. *TDNetGen* provides two examples of custom exporters, provided in the *custom_data* folder. The first is the power flow program ARTERE developed at the University of Liege and available at. Compared to MATPOWER, this software includes the modelling and treatment of OLTC transformers and is computationally more efficient. The second custom exporter is for the academic time-domain, dynamic simulation software RAMSES . In order to execute a dynamic simulation, the dynamic data for both the TN generators and controllers as well as the DGs, are required. For the TN, the dynamic data are taken from . For the DNs, the DGs in feeders 1-4 are modelled as small synchronous machines following , while the DGs in feeders 5-8 use the distributed PV system model (PVD) . The initialization of the dynamic models takes place in RAMSES based on the operating point provided by *TDNetGen*. Using the above two examples, users can create their own exporters to a variety of software packages. One of the future developments will be the implementation of an exporter to CIM format , to facilitate the integration of the models to other software.
# Example Test Cases {#sec:Power-Flow-Test-Cases}
In this section, we present some combined T&D test systems generated by *TDNetGen* focusing on producing some operationally challenging test cases.
## Power-flow scenarios
One of the main problems associated with increased penetration of DGs and in particular RES are possible over-voltage problems arising at high generation levels. *TDNetGen* is able to generate such scenarios by increasing the penetration level. As the penetration level increases, more buses experience over-voltages, in the DNs but also in the TN. Figure [\[fig:over\]](#fig:over){reference-type="ref" reference="fig:over"}, shows the number of TN buses with over-voltages (above 1.15 pu) as the penetration level increases. In these scenarios, the operating point produced by *TDNetGen* is not optimised (*Run OPF='false'*). If an OPF solution is selected, then the operating point can be corrected by re-dispatching the TN generators and the DGs in feeders 1-4 of each DN. Figures [\[fig:pl0_20\]](#fig:pl0_20){reference-type="ref" reference="fig:pl0_20"} and [\[fig:pl25_75\]](#fig:pl25_75){reference-type="ref" reference="fig:pl25_75"} show the voltage profiles at all the buses located in a specific DN for different penetration levels. In these figures, the 75 DN buses are numbered consecutively starting from feeder 1 up to 8 and from the bus closest to the DN substation to the one furthest (see Fig. [\[fig:one_line_dn\]](#fig:one_line_dn){reference-type="ref" reference="fig:one_line_dn"}). The continuous curve represents the voltage profile of the DN without any DG and where the voltage set-point of the OLTC is at 1.03 pu[^1]. It can be seen that as the penetration level increases, the voltage profile of the feeder is changing. As the penetration level increases, the total power demand of the DNs from the TN decreases. In cases of higher than 100% penetration, when the DGs cover the local load consumption as well as the DN losses, then a reverse power flow arises. In these extreme cases, the power flows from the DNs to the TN and it can cause voltage violations and challenge the effectiveness of existing management, control, and protection schemes. Figure [\[fig:rev_pf\]](#fig:rev_pf){reference-type="ref" reference="fig:rev_pf"} shows the total power transferred from the TN to all the DNs for different penetration levels. It can be seen that at about 115%, the DNs cover their own local consumption and losses and there is no power exchange with the TN. It should be noted that individual DNs might reach a reverse power flow condition at lower or higher penetration levels due to their different operating points.
## A dynamic simulation scenario using RAMSES
As explained in , custom exporters can be used to translate the data into other formats used in other software than MATPOWER. One of the example exporters with the toolbox produces the data to be used in RAMSES, an academic, time-domain, dynamic simulation software . This simulator uses a topologically-based, domain decomposition method to partition the system into the TN and the multiple DNs. Then, it solves the sub-problems defined over each sub-domain at each discrete time instance, while treating the interface variables with a Schur-complement approach. Parallel computing techniques are employed to accelerate the system simulation. A T&D model was generated with the parameters *Constant load='true'*, *Large system='false'*, *Run OPF='false'*, *Penetration level=50%*, *Oversize=2.0*. The data was exported into RAMSES format, leading to a system with 141 DNs in the Central area, with a total of 10782 buses, 20 large (connected to the TN) and 564 small synchronous generators (connected to the DN feeders 1-4), 2538 distributed PV systems (connected to the DN feeders 5-8), and 10595 voltage-dependent loads. The considered disturbance is a 5-cycle (100 ms) short-circuit near bus 4032, cleared by opening line 4032-4042. The response is simulated over a horizon of 250 s with a time-step of 1 cycle (20 ms). This scenario leads to a long-term voltage collapse driven by the load-power restoration caused by the OLTCs restoring the DN voltages. Figure [\[fig:ramsesTN\]](#fig:ramsesTN){reference-type="ref" reference="fig:ramsesTN"} shows the evolution of the TN voltages while Fig. [\[fig:ramsesDN\]](#fig:ramsesDN){reference-type="ref" reference="fig:ramsesDN"} shows the DN voltages at the buses in a DN attached to the TN bus 1041. More information on the analysis, instability detection and corrective control of such unstable scenarios in combined T&D systems is given in.
## Computational cost of *TDNetGen*
Depending on the choice of parameters, the run-time to generate the T&D system can vary significantly. However, this is a cost occurred only once to generate the test system, which can then be used independently of *TDNetGen*. The most computationally intensive task is the OPF solution of the system to optimise the operating point (if selected). Table [1](#tab:cpu){reference-type="ref" reference="tab:cpu"} shows the average execution times to generate a T&D system[^2] as well as the average number of times the OLTC setpoints are updated during an execution. The overall execution time does not exceed 4 minutes, using a standard laptop computer.
:::
# Conclusion {#sec:conclusion}
In this paper, an open-source MATLAB toolbox was presented, that is able to generate synthetic large-scale T&D network test cases. These models can be used to develop or test solutions targeting problems specific to combined T&D systems (optimisation, market clearing, system security, ancillary services by DNs, etc.). The toolbox is based on the widely used MATPOWER software and uses the well-known Nordic system as the TN upon which the combined T&D system is built. *TDNetGen* is highly parametrizable, allowing to generate models with a variety of characteristics, replicating common problems in systems with high penetration of RES. In addition, medium-(\(\sim 10000\) buses) to large-scale (\(\sim 22000\) buses) test-system models can be produced, allowing to test the scalability of algorithms proposed by the users. Moreover, the model data can be shared among researchers to provide a common testing platform. Finally, the open-source nature of the toolbox allows the users to expand the code, adding new functionality, or building custom export functions to import the generated models into their own software. The latest version of the toolbox can be found at. | {'timestamp': '2017-11-15T02:08:08', 'yymm': '1706', 'arxiv_id': '1706.01656', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01656'} |
null | null |
# Introduction {#intro}
When a polycrystalline material is deformed, its microstructure generally experiences a reorientation of the crystal lattices of each grain towards a preferential distribution of orientations known as *crystallographic texture*. The study of texture evolution is important because textured metals typically exhibit plastic anisotropy, which plays a significant role on mechanical properties. Predicting the evolution of deformation-induced texture and the accompanying plastic anisotropy is the subject of polycrystal plasticity models. These models are typically formulated assuming that the microstructure of the polycrystal is associated with a representation of microscopic crystals whose individual responses, on average, determine the macroscopic response of the polycrystal. At the level of each grain, plastic deformation occurs by the standard mechanism of dislocation slip, and so (i) constitutive equations that relate dislocation motion to crystal deformation must be defined, and (ii) an averaging scheme that relates the response of individual crystals to the macroscopic stress-strain response of the polycrystal must also be defined. For single crystals, a multiplicative kinematic decomposition of the deformation gradient into elastic and plastic parts is typically used. This decomposition adequately describes the distinctly different kinematical mechanisms that operate during the plastic deformation of a crystal. It was formally introduced in continuum plasticity, and then applied to describe the kinematics of single crystals. A feature of this decomposition is that it introduces an intermediate configuration between the reference and current configurations which is obtained by unloading the crystal to a stress-free state. The elasto-viscoplastic constitutive equations are generally written relative to this relaxed configuration. Many numerical procedures have been proposed to integrate the crystal constitutive equations, generally implicit and semi-implicit procedures which are developed differently by particular selection of the primary variables (stresses, shear rates, plastic deformation gradient, etc.). Polycrystal plasticity models appear in various levels of sophistication. Along the venerable Sachs and Taylor models--in which the aggregate deformation or stress is computed by averaging from the individual crystal values--, self-consistent models have been developed and applied that express the global deformation in terms of linearized viscoplastic moduli that must be adjusted self-consistently. Models that spatially resolve grain boundaries (GB) have started to gain traction recently thanks to a higher efficiency of numerical solvers and a wider availability of computational resources. Roters *et al.* have provided a comprehensive review of the different variants of such approaches, which enable the calculation of the fine spatial features of strain and stress fields, including grain shape changes and nonuniform deformation. Some of these advances have also been discussed by Knezevic *et al.*. However, in the above models, grain boundary processes--which are known to be relevant at high stresses and temperatures--cannot be captured by construction. For example, fundamental grain boundary properties such as energies and mobilities are extraneous to spatially-resolved *standard* (poly)crystal plasticity models. The aim of this paper is to present a framework that preserves the ability to model intra-grain plasticity, while at the same time enabling a straightforward generalization to include grain boundary processes. To this end, we develop a 'diffuse'-interface crystal plasticity model for polycrystalline materials based on a representation of grain boundaries as a special subclass of geometrically necessary dislocations (in the sense defined by Cermelli and Gurtin ). In this model, with single crystal as the reference configuration, a stress-free polycrystal is constructed by imposing a piecewise constant rotation field and its transpose as the lattice and plastic distortions respectively. To make the resulting model numerically tractable, we regularize the piecewise constant rotation field, resulting in a diffuse interface model, that preserves the zero-stress character of the grain boundaries. Our main intent here is to introduce the model and its potential, and perform a verification exercise before launching into more ambitious undertakings where grain boundary phenomena can be properly modeled. In the following sections we lay out the essential theoretical developments of our model and provide a verification exercise of the numerical implementation.
# Classical crystal plasticity for single crystals
For reference, in this section, we introduce the framework of crystal plasticity for single crystals as a starting point. A body is represented as an open subset \(\mathcal B\) of the three-dimensional Euclidean space \(\mathbb R^3\). Let \(\mathcal B^0 \subset \mathbb R^3\) represent the reference configuration of the body. The position of an arbitrary material point in the reference configuration is denoted by \(\bm X\). A time-dependent deformation map is given by a one-to-one function \(\bm y(\bm X,t)\), such that \(\det \bm F \ne 0\), where \[\begin{aligned}
\bm F(\bm X,t) := \nabla \bm y(\bm X,t) \end{aligned}\] is the gradient of the deformation map. In the theory of crystal plasticity, there exists a decomposition of the deformation gradient given by \[\begin{aligned}
\bm F = \bm F^{\rm L} \bm F^{\rm P}, \label{eqn:fefp} \end{aligned}\] where \(\bm F^{\rm L}\) and \(\bm F^{\rm P}\) are lattice[^1] and plastic components of \(\bm F\) respectively, and \(\det \bm F^{\rm P}=1\). In this paper, \(\bm F^{\rm P}\) represents the deformation gradient of an infinitesimal material element, attributed to dislocation slip through its volume. Since such a process renders the lattice invariant, it follows that \(\bm F^{\rm P}\) leaves the lattice undeformed. \(\bm F^{\rm L}\) represents the deformation of the material due to the deformation of its underlying lattice. Note that \(\bm F^{\rm L}(\bm X,t)\) and \(\bm F^{\rm P}(\bm X,t)\) *need not* be gradients of a deformation map. Instead, since \(\bm F^{\rm L}\) and \(\bm F^{\rm P}\) are invertible, they represent deformation of an infinitesimally small neighborhood of \(\bm X\) at time \(t\). In other words, \(\bm F^{\rm P} \mathrm d \bm X\) represents the deformation of a differential material element \(d\bm X\). The collection of all deformed differential material elements is referred to as *lattice configuration*. In this sense, \(\bm F^{\rm P}\) maps the reference configuration to the lattice configuration, and \(\bm F^{\rm L}\) maps the lattice configuration to the deformed configuration. As is customary, dislocations move on slip systems \(\alpha=1,2,\dots,A\), where each \(\alpha\) defines a glide direction \(\bm s^\alpha\) and a slip plane normal to \(\bm m^\alpha\). These two are vectors in the lattice configurations such that \[\begin{aligned}
|\bm s^\alpha| = |\bm m^\alpha| = 1; \quad \bm s^\alpha \cdot \bm m^\alpha = 0; \quad \bm s^\alpha, \bm m^\alpha = \text{constant}. \end{aligned}\] Evolution of \(\bm F^{\rm P}\) is governed by slip rates \(v^\alpha(\bm X,t)\) on individual slip systems via the flow rule \[\begin{aligned}
\label{eqn:flow} \dot \bm F^{\rm P} = \bm L^{\rm P} \bm F^{\rm P}, \end{aligned}\] where \[\begin{aligned}
\bm L^{\rm P}(\bm X,t) := \sum_{\alpha=1}^A v^\alpha(\bm X,t) \bm s^\alpha \otimes \bm m^\alpha. \label{eqn:defn_Lp} \end{aligned}\] If the free energy density, denoted by \(\psi\), depends on the lattice Lagrangian strain \[\begin{aligned}
\bm{\mathbb E}^{\rm L}:=((\bm F^{\rm L})^{\rm T} \bm F^{\rm L}-\bm I)/2, \label{eqn:lag_strain} \end{aligned}\] then the evolution equations of crystal plasticity are given by the flow rule in , along with the following macroscopic and microscopic force balance equations:
- Macroscopic force balance
::: {.subequations}
\[\begin{aligned}
\Div \bm P(\bm X,t) &= \bm 0, \quad \bm X \in \mathcal B^0, t>0,\\ \bm u &= \bm u^0 \text{ on } \partial \mathcal B^0, \end{aligned}\] [\[eqn:macroscopic\]]{#eqn:macroscopic label="eqn:macroscopic"}
:::
where \[\begin{aligned}
\bm P := \bm F^{\rm L} \psi,_{\bm{\mathbb E}^{\rm L}} (\bm F^{\rm{P}})^{\rm{-T}}, \end{aligned}\] is the first Piola--Kirchhoff stress tensor, and \(\psi,_{\bm{\mathbb E}^{\rm L}}\) denotes the derivative of \(\psi\) with respect to \(\bm{\mathbb E}^{\rm L}\).
- Microscopic force balance for each slip system \(\alpha\) \[\begin{aligned}
\label{eqn:microscopic} b^\alpha v^\alpha(\bm X,t) = \psi,_{\bm{\mathbb E}^{\rm L}} \bm m^\alpha \cdot \bm C^{\rm L} \bm s^\alpha, \end{aligned}\] where \(b^\alpha \ge 0\) is the inverse of the mobility associated with the slip \(v^\alpha\), and \(\bm C^{\rm L}=(\bm F^{\rm L})^{\rm T} \bm F^{\rm L}\) is the right Cauchy-Green strain tensor.
The non-negativity of the inverse mobilities is a necessary condition for thermodynamic consistency. The expression on the right-hand-side of is commonly referred to as the resolved shear stress. See the work by Gurtin for a thermodynamically consistent derivation of and . In standard crystal plasticity, a stress-free single crystal at \(t=0\) is modeled using the initial conditions \[\begin{aligned}
\bm F^{\rm P}(\bm X,0) = \bm F^{\rm L}(\bm X,0) \equiv \bm I. \label{eqn:initial} \end{aligned}\] Note that, the above initial conditions are also used for polycrystals, with the difference that \(\bm L^P\) is evolved in a piecewise way in each grain due to the different orientation of the slip systems, and the free energy density given by \(\psi(\bm R^{\rm T} \bm{\mathbb E}^{\rm L} \bm R)\), where \(\bm R\) is a piecewise constant rotation field describing the initial orientation of grains.[^2] In the next section, we first present a diffuse-interface polycrystal plasticity model which operates at a length scale where all grain boundaries are resolved explicitly. In contrast with assumption [\[eqn:initial\]](#eqn:initial){reference-type="eqref" reference="eqn:initial"}, the proposed framework gives us access to grain boundary dislocation densities, thus enabling us to model grain boundary energies.
# Polycrystal plasticity
Consider a sharp-interface polycrystal, *i.e.* one where the orientation of the lattice is constant in the interior of one grain and has a jump discontinuity along the grain boundary. In this context, crystal plasticity is studied by having the stress-free polycrystal as the reference configuration. Due to the variation in orientation of the grains, the elastic and plastic response of each grain is different. Therefore, the elastic moduli and the slip systems (\(\bm s^\alpha\) and \(\bm m^\alpha\)) are piecewise constant, with jump discontinuities along the grain boundaries. If the polycrystal is stress-free at \(t=0\), then the initial conditions are identical to . Thus, within this framework, polycrystal plasticity is identical to single crystal plasticity with the caveat that the elastic moduli, \(\bm s^\alpha\) and \(\bm m^\alpha\) are piecewise constant. While this model is remarkably simple, it is not straightforward to generalize it to model grain boundary-mediated deformation, such as shear-induced grain boundary motion, grain shrinkage and rotation, grain boundary sliding, etc. These phenomena can become important during plastic deformation at high stresses and/or temperatures, such as during recovery, recrystallization, and grain growth. In the following section, we present an alternate framework that lays the foundation to model polycrystal plasticity with grain boundary evolution.
## Diffuse-interface polycrystal plasticity
The success of single crystal plasticity in describing the materials deformation lies in precisely identifying the *independent* mechanisms involved, and attributing them appropriately to the evolution of \(\bm F^{\rm P}\). For example, the rate of change is \(\bm F^{\rm P}\) due to dislocation slip is identified with the slip rate projected on each slip system by way of the Schmid tensor. Similarly, additional mechanisms such as dislocation climb are built into the evolution law for \(\bm F^{\rm P}\). In addition to dislocations, a grain boundary sweeping through a material also results in plastic distortion. For example, consider a circular grain with lattice orientation \(\theta_2\) embedded in a larger grain with orientation \(\theta_1\). The misorientation of \(|\theta_2-\theta_1|\) results in a grain boundary energy. In order minimize the internal energy, the circular grain shrinks. As the circular grain boundary sweeps through the material, the lattice in the swept region rotates from an initial configuration of \(\theta_1\) to \(\theta_2\), while the rest of the lattice remains unchanged. If \(\bm F^{\rm P}\) is equal to identity during this process, then this results in an incompatible \(\bm F\). This conclusively suggests that \(\bm F^{\rm P} \not \equiv \bm I\) in the swept area. In other words, grain boundary motion always results in plastic distortion. Therefore, in the spirit of modeling plasticity due to bulk dislocations, plasticity due to grain boundary motion may thus be modeled by identifying the mechanism for the accompanying plastic distortion, and include it in the evolution law for \(\bm F^{\rm P}\). Identifying the pertinent GB-mediated plastic mechanisms is highly non-trivial. For example, recent atomistic simulations have revealed that for certain misorientations the interior grain not only shrinks but also rotates with *no* dislocation activity in the bulk. This suggests that unlike dislocation slip, there is no unique fundamental evolution law for \(\bm F^{\rm P}\) that can be attributed to the motion of a grain boundary with a given misorientation. Therefore, we take an alternate approach to modeling plasticity due to grain boundary motion. The central idea behind this approach is to identify dislocations as the basic defect carriers, and build grain boundaries as continuum aggregates of dislocations. Therefore, any motion of grain boundary is viewed as a collective motion of dislocations that form the boundary. The most important advantage of this approach is plastic distortion due to grain boundary motion emerges from the original flow rule given in without identifying any new mechanisms. This approach can model phenomena such as shear-induced grain boundary motion, grain boundary sliding and grain rotation. We next build a framework of polycrystal plasticity based on the idea described above. [\[sec:diffuse\]]{#sec:diffuse label="sec:diffuse"} Let \(\bm R^0(\bm X) \in SO(3)\), a step function in the space of special orthogonal tensor fields, represent the lattice rotation field in the polycrystal, with piecewise-constant values in each grain and smooth transitions across grain boundaries. In contrast to [\[eqn:initial\]](#eqn:initial){reference-type="eqref" reference="eqn:initial"}, the initial state of the polycrystal is chosen to be: \[\begin{aligned}
\bm F^{\rm L}(\bm X,0) = \bm R^0(\bm X), && \bm F^{\rm P}(\bm X,0) = \bm R^0(\bm X)^{\rm T}, \label{initFeFp} \end{aligned}\] resulting in \[\begin{aligned}
\bm F(\bm X,0) \equiv \bm I. \label{initF} \end{aligned}\] The decomposition given in is the central idea of the current framework, and we now describe its physical significance. demonstrates the decomposition given in for the construction of a grain boundary in a bicrystal. Recall that \(\bm F^{\rm P}\) deforms the material leaving the lattice fixed as shown in . On the other hand, \(\bm F^{\rm L}\) deforms the lattice resulting in a total deformation gradient \(\bm F\) that is compatible. Comparing the reference and the final configurations, seems contradictory since the material is shown to be deformed although \(\bm F \equiv \bm I\). We now discuss the correct mathematical interpretation that resolves this contradiction. We begin by noting that \(\bm F^{\rm P}(\bm X,0) = \bm R^0(\bm X)^{\rm T}\) qualifies to be a plastic distortion due to dislocation slip, since a rotation can always be expressed as a product of three shear deformation tensors.[^3] Interpreting the three resulting shear deformations as lattice-invariant shears obtained due to dislocation slips, the rotation tensor \(\bm F^{\rm P}(\bm X,0)\) may be interpreted as a lattice-invariant deformation. Since an arbitrary rotation rotates the material, it may seem contradictory for it to leave the lattice invariant (except of course when the rotation belongs to the point group of the lattice). The correct mathematical interpretation of a "lattice-invariant" rotation is given using the notion of weak-convergence discussed in Appendix [\[appendixFeFp\]](#appendixFeFp){reference-type="ref" reference="appendixFeFp"}. In short, weak convergence represents convergence of functions/distributions on the "average". In Appendix [\[appendixFeFp\]](#appendixFeFp){reference-type="ref" reference="appendixFeFp"}, we show that, for a sequence of lattice constants \(a^i\to 0\) (as \(i\to\infty\)), \(\bm F^{\rm P}(\bm X,0)\) has to be viewed as a *weak-limit* of a sequence of deformations \((\bm F^{\rm P})^i\) that leave the \(a^i\)-lattice invariant. Therefore, interpreting \(\bm F^{\rm P}(\bm X,0)=\bm R^{\rm T}(\bm X)\) and \(\bm F \equiv \bm I\) for a discrete lattice in an average sense resolves the apparent contradiction described in the previous paragraph. An important consequence of the decomposition given in is that the resulting polycrystal is stress-free since the Lagrangian strain, defined in , is equal to zero. Therefore, eq. [\[initFeFp\]](#initFeFp){reference-type="eqref" reference="initFeFp"} describes a polycrystalline state which is obtained from a reference single crystal by the right amount of slip in each grain such that grains undergo relative rotation but the polycrystal remains stress free. An advantage of the above construction is that we have immediate access to the grain boundary dislocation density content in the form of the *geometrically necessary dislocation density* \(\bm G\) tensor defined as \[\begin{aligned}
\bm G = \bm F^{\rm P} \Curl \bm F^{\rm P}, \label{eqn:G} \end{aligned}\] where \(\Curl\) denotes the curl of a tensor field with respect to the material/reference coordinate.[^4] For a given normal \(\bm n\) in the lattice configuration, the vector \(\bm G^{\rm T}\bm n\) measures the net Burgers vector of dislocation lines per unit area passing through a plane of normal \(\bm n\). Using the decomposition of \(\bm F\) discussed above, we can now, in principle, study a polycrystal under a single boundary-value problem. Numerically, the problem still does not enjoy the nice characteristics of its single crystal counterpart as \(\bm F^{\rm L}\) and \(\bm F^{\rm P}\) are discontinuous. In order to overcome this challenge, we introduce a smooth-interface version of the above sharp-interface model. This can be achieved by constructing a stress-free diffuse interface crystal plasticity at \(t=0\) with \(\bm F^{\rm P}\) a smoothened step function in the space of rotation fields. This alteration ensures that all the resulting fields are smooth.
# Numerical implementation
In this section, we discuss a three-dimensional numerical implementation of tensile tests of polycrystals of varying textures using the diffuse-interface model introduced in . The main aim of this section is to demonstrate the robustness of the diffuse-interface model. We implement a simpler version of a crystal plasticity model for body-centered cubic (bcc) Fe used by Barton, Arsenlis, and Marian that incorporates the role of latent hardening into the mobility variable in . The microscopic force balance we use in this implementation is given by \[\begin{aligned}
v^\alpha(\bm X,t) = v^\alpha_0 \left ( \frac{\tau^\alpha}{g^\alpha}\right )^{1/m}, \label{eqn:microscopic_mecking} \end{aligned}\] where \(v^\alpha_0\) the references shear rate, \(g^\alpha(\bm X,t)\) is the slip system strength that captures the operating hardening mechanism, \(\tau^\alpha(\bm X,t)=\psi,_{\bm{\mathbb E}^{\rm L}} \bm m^\alpha \cdot \bm C^{\rm L} \bm s^\alpha\) is the resolved shear stress, and \(m=0.05\) is the strain-rate sensitivity exponent. The slip strength \(g^\alpha\) depends on the network dislocation density \(\rho_n(\bm X,t)\) via Taylor hardening: \[\begin{aligned}
g^\alpha = g_0 + b\mu_0\sqrt{h_n \rho_n}, \label{eqn:g} \end{aligned}\] where the constant \(g_0=90\) MPa refers to the slip strength in a single crystal, \(\mu_0=86\) GPa is the rigidity modulus of iron, and \(h_n=0.125\). The network dislocation density \(\rho_n\) in evolves according to the Kocks--Mecking type evolution model: \[\begin{aligned}
\dot \rho_n = v (k_1 \sqrt{\rho_0 \rho_n}-k_2 \rho_n), \label{eqn:rho} \end{aligned}\] with \[\begin{aligned}
k_2 = k_{20} \left ( \frac{v_{k0}}{v} \right )^{\frac{1}{n}}. \end{aligned}\] The variable \(v = \sum_\alpha |v^\alpha|\) is the aggregate slip rate, \(\rho_0=10^{12}\) m\(^{-2}\) is the reference network bulk dislocation density, and the Kocks-Mecking parameters \(k_1\), \(k_{20}\), and \(v_{k0}\) are equal to \(450\), \(14\) and \(10^{10}\) s\(^{-1}\) respectively. Finally, the elastic free energy \(\psi\) is taken to be of the form: \[\psi(\bm{\mathbb E}^{\rm L})=\frac{1}{2}\mathbb{C}~\bm{\mathbb E}^{\rm L}\cdot\bm{\mathbb E}^{\rm L}\] where \(\mathbb{C}\) is the elasticity matrix, which for a cubic material is fully characterized by three independent elastic constants whose values for Fe are: \(C_{11}=228\) MPa, \(C_{12}=132\) MPa, and \(C_{44}=116\) MPa.
## Finite element implementation
The finite element method is used to solve the resulting system of equations in , , and , with the displacement field \(\bm u\), the plastic distortion \(\bm F^{\rm P}\), and the bulk network dislocation density \(\rho_n\) as unknowns. In particular, the three displacement variables \(u_1\), \(u_2\) and \(u_3\) are interpolated using the Lagrange quadratic finite elements, while \(\rho_n\) is interpolated using the Lagrange linear finite elements. Recall that at \(t=0\), \(\bm F^{\rm P}\) is a field in \(SO(3)\). This implies that it satisfies the condition of orthogonality, i.e. \((\bm F^{\rm{P}})^{\rm T} \bm F^{\rm P}(\bm X,0) \equiv \bm I\). On the other hand, the components of \(\bm F^{\rm P}\) interpolated using the Lagrange finite elements cannot satisfy the orthogonality constraint in the interior of the finite elements. Therefore, the components of \(\bm F^{\rm P}\) cannot be interpolated using the Lagrange finite elements. Instead, using the polar decomposition, \(\bm F^{\rm P}\) is expressed as \(\bm R^{\rm P} \bm U^{\rm P}\), where \(\bm R^{\rm P}(\bm X,t) \in SO(3)\), and \(\bm U^{\rm P}(\bm X,t)\) is the resulting positive-definite plastic stretch tensor. Using the angle-axis representation for rotation tensors, \(\bm R^{\rm P}\) is expressed in terms of a vector \(\bm q \in \mathbb R^3\): \[\begin{aligned}
\bm R(\bm q) = \bm I + \frac{\sin |\bm q|}{|\bm q|} \bm W + \frac{1}{2} \left [ \frac{\sin(|\bm q|/2)}{(|\bm q|/2)} \right ]^2 \bm W^2, \end{aligned}\] where \(\bm W\) is the skew-symmetric matrix associated with \(\bm q\), and \(|\bm q|\) and \(\bm q/|\bm q|\) represent the angle and axis of the rotation tensor. Lagrange linear finite element interpolation is then chosen for the variables \(\bm U^{\rm P}\) and \(\bm q\), from which \(\bm F^P\) is locally computed as \(\bm F^P=\bm R^P(\bm q)*\bm U^P\). This method guarantees that \(\bm F^P(\bm X,0)\) can describe an exact plastic rotation field without numerical artifacts due to the interpolation method. To simulate tensile tests of polycrystals with different textures, we impose the boundary conditions shown in . The initial conditions for \(\bm u\), \(\bm U^{\rm P}\) and \(\rho_n\) are chosen to be \[\begin{aligned}
\bm u(\partial \mathcal B^0) = \bm 0, \quad \bm U^{\rm P}(\mathcal B^0) \equiv \bm I,\text{ and} \quad \rho_n(\mathcal B^0) \equiv 20 \rho_0, \label{eqn:initial_conditions} \end{aligned}\] respectively. The remaining initial conditions for \(\bm q\), which defines the texture of the polycrystal, is discussed in . The system of equations , , and are evolved in a segregated manner using the MUMPS direct solver, and BDF (Backward Differential Formula) time stepping algorithm implemented in `COMSOL5.2`. In particular, due to the highly nonlinear nature of expressed in \(\bm q\) and \(\bm U\), we enable the "automatic highly nonlinear (Newton)" option to obtain well-behaved solutions. On the other hand, we rely on the default "constant (Newton)" option for solving , and . All simulations were performed on a finite element mesh with \(99883\) elements, and \(595620\) degrees of freedom.
## Construction of polycrystals with different textures {#sec:polycrystal}
In this section we describe the generation of diffuse interface polycrystals of different textures. The grain orientations are outputted in the form of a smoothened rotation vector field \(\bm q(\bm X,0)\) which serves as an initial condition along with those given in . A stress-free polycrystal with \(N\) grains is constructed by randomly choosing \(N\) points, \(\bm{\mathcal P}^1,\dots, \bm{\mathcal P}^N\), within the domain, and constructing a corresponding *diffuse* Voronoi tessellation. The grain orientations are prescribed by associating random rotation vectors \(\bm q^1,\dots,\bm q^N\) to each grain. The diffuse tessellation is constructed using a grid of size \(100\times100\times100\), and assigning each grid point to a grain based on the Voronoi construction, i.e. a grid point \(p_i\) is associated with a grain \(\alpha\) if \[\begin{aligned}
\dist(\bm p^i,\bm{\mathcal P}^\alpha) < \dist(\bm p^i, \bm{\mathcal P}^\beta), \quad \forall \beta \ne \alpha, \label{eqn:voronoi} \end{aligned}\] where \(\dist(\bm p^i,\bm{\mathcal P}^\beta)\) is the distance between \(\bm p^i\) and \(\bm{\mathcal P}^\beta\). Finally, the rotation vector \(\bm q^\alpha\) is associated to the grid point \(\bm p^i\). The polycrystal is outputted in the form of the rotation vector field on the grid which is then interpolated as a smooth vector field \(\bm q(\bm X,0)\) using the *nearest neighbor* interpolation implemented in `COMSOL5.2`. Therefore, the texture of the resulting collection of grains depends on the distribution of the initial collection of \(N\) random points, and the grain boundary "thickness" is inversely proportional to the resolution of the grid. The pseudocode for the above algorithm is described in Algorithm [\[alg\]](#alg){reference-type="ref" reference="alg"}. We study textures with (i) a log normal distribution of grain sizes[^5], (ii) elongated grains, and (iii) flat grains. The size of a grain \(\alpha\) is defined as \[\begin{aligned}
{\rm size}(\alpha) = \min_\beta\{\dist(\bm{\mathcal P}^\alpha,\bm{\mathcal P}^\beta): \beta \in \{1,\dots,N\}, \beta \ne \alpha\}. \label{eqn:size} \end{aligned}\] Grains with a log normal distribution of sizes are generated by sampling the initial \(N\) points from a log normal distribution based on Algorithm [\[alg\]](#alg){reference-type="ref" reference="alg"} described in Appendix [\[appendix\]](#appendix){reference-type="ref" reference="appendix"}. We use the standard Euclidean metric for \(\dist\) in in the generation of the texture with log normal distribution of grain sizes. On the other hand, elongated and flat grains with aspect ratio equal to 4 are generated by sampling the initial \(N\) points from a Dirac probability measure supported on \(0.2L\), and using the metric \[\begin{aligned}
\dist(\bm x,\bm y) &= \left (\frac{x_1-y_1}{sx}\right )^2+ \left (\frac{x_2-y_2}{sy}\right )^2+ \left (\frac{x_3-y_3}{sz}\right )^2, \end{aligned}\] with the scales \(sx=1\), \(sy=1\), \(sz=4\) for elongated grains, and \(sx=0.25\), \(sy=0.25\), and \(sz=1\) for flat grains. Polycrystals with the three textures studied in this paper are shown in , with the colors obtained by plotting the \(q_3\) component, indicating different grains.
# Results {#sec:results}
In this section, we present our results of the simulated tensile tests on polycrystals of varying textures. Figure [\[fig:logNormal_dd\]](#fig:logNormal_dd){reference-type="ref" reference="fig:logNormal_dd"} shows a color plot of the grain boundary dislocation density for a polycrystal with log normal grain size distribution, calculated using . Note that the field \(\bm G\) is not available in the classical polycrystal plasticity implementation. In the proposed model, the initialization [\[initFeFp\]](#initFeFp){reference-type="eqref" reference="initFeFp"} allows to construct a kinematically consistent grain boundary structure which evolves in time as a consequence of slip in each grain. shows plots of the intermediate stress \(\psi,_{\bm{\mathbb E}^{\rm L}}\) versus the first axial component of the total strain \(\bm E:=(\bm F^{\rm T} \bm F-\bm I)/2\) for different textures and loading orientations. In addition, also shows the variation of the normalized dislocation density \(h\) with respect to the total strain. We have verified that the plots shown in are insensitive to further mesh refinement. In addition, since the computed properties are aggregates, as expected, we ensured that the results are not sensitive to grain boundary thickness. We expect that local properties such as stress concentration will be sensitive to the choice of grain boundary thickness. We also compute the Taylor factor, which is known to be \(2.9\) for an equiaxed bcc random polycrystal. The Taylor factor \(M\) is defined as the ratio of the aggregate microscopic shear rate in a polycrystal to the macroscopic shear rate. It is defined using the following equivalence of the power supplied by external loads to the power dissipated due to slip: \[\begin{aligned}
\bm P \cdot \dot{\bm F} = \sum_\alpha \tau^\alpha v^\alpha. \label{eqn:equiv} \end{aligned}\] Assuming there exists a constant critical resolved shear stress \(\tau^{\rm c}>0\) for every slip system at the which a crystal slips, can be simplified to \[\begin{aligned}
\bm P \cdot \dot{\bm F} = \tau^{\rm c} \sum_\alpha |v^\alpha|. \label{eqn:equiv_tau} \end{aligned}\] The Taylor factor \(M\) is defined as
where we have used to arrive at the last equality, and \(\langle \cdot \rangle\) denotes spatial average. The traditional definition of the Taylor factor given in cannot be used in a straightforward manner in our implementation since the slip does not occur precisely at a critical load. In fact, when implemented, and neither agree, nor converge with time. On the other hand, by factoring out \(\langle \sum_\alpha \tau^\alpha\rangle\) instead of \(\tau^{\rm c}\) in , we show that the following two definitions for \(M\) given by
are not only consistent with each other, but also converge to a constant value as shown in . The converged values of the Taylor factors for different textures are listed in Table [1](#table:taylor){reference-type="ref" reference="table:taylor"}.
# Discussion and conclusions
Most metals and alloys in usable form display an internal microstructure characterized by a collection of grains with different lattice orientation separated by grain boundaries. Metals deformation, particularly at high temperatures and stresses, such as during hot working, involves not just intragranular plasticity but also plasticity controlled by grain boundary mechanisms. Standard formulations of crystal plasticity decouple both types of deformation, probably due to our good deal of understanding about low temperature processes, *e.g.* cold working, which tends to dominate our thinking of plasticity. Indeed, this decoupling has been the governing principle behind the development of new methodologies to study recrystallization in metals. However, during dynamic phenomena such as *continuous* dynamic recrystallization, bulk and grain-boundary plastic processes can occur simultaneously, and therefore the underlying plasticity model must be capable of capturing both types of deformations concurrently. This is the motivation behind the present work: to devise a computational model that combines bulk and grain boundary plasticity by design within the same framework. Our purpose at the moment is simply to demonstrate that our formulation is capable of rendering the same response as standard crystal plasticity models for conventional problems in polycrystal plasticity. Only after fulfilling this step can we truly apply our methodology to phenomena involving grain boundary processes. We have undertaken this verification exercise by solving the same problem, standard Taylor hardening in body-centered cubic Fe, using both methodologies, and comparing the results obtained. To explore the capabilities of our model further, we have considered several different textures and misorientation ranges and have calculated the associated Taylor factors. In all cases, our results agree with those obtained using standard polycrystal plasticity. In summary, we have developed a diffuse-interface model for polycrystalline materials deformation that expresses grain boundaries as a special class of geometrically necessary dislocations, such that the stress-free nature of the polycrystalline structures obtained is naturally recovered. We have tested the robustness of the method by simulating tensile tests and calculating Taylor factors for polycrystals of varying textures. Our model provides a pathway from which grain boundary energies and mobilities can eventually be obtained directly from dislocation densities, which opens the door to integrated models of intragranular and grain boundary-governed plasticity such as recrystallization in hot working.
**Authors' Contributions**-NCA developed the theory, tested the model, ran the simulations, and wrote most of the manuscript. GP assisted with the theoretical developments and with the solution procedure. JM contributed to the theoretical developments and to the writing of the paper. All authors read and approved the final manuscript.\
**Competing Interests**-The authors declare that they have no competing interests.
[^1]: In the literature, it is more common to refer to the lattice distortion as an elastic distortion using the notation \(\bm F^{\rm E}\). Since the decomposition given in is purely geometric in nature (as opposed to energetic), we prefer the term "lattice distortion" denoted by \(\bm F^{\rm L}\), a terminology adopted by Clayton.
[^2]: It is important to note that, within the framework of crystal plasticity, a constitutive response function of the form \(\psi(\bm R^{\rm T} \bm{\mathbb E}^{\rm L} \bm R)\) with a non-constant \(\bm R\) *does not* imply \(\bm F^{\rm L}=\bm R\) and \(\bm F^{\rm P} \equiv \bm I\) as this would result in an incompatible \(\bm F\).
[^3]: For example, a rotation by angle \(\theta\) about the \(z\)-axis can be decomposed multiplicatively into three shear deformations as \[\begin{aligned}
\begin{bmatrix} \cos \theta &-\sin\theta \\ \sin \theta & \cos \theta \end{bmatrix} = \begin{bmatrix} 1 & -\tan(\frac{\theta}{2}) \\ 0 & 1 \end{bmatrix} \begin{bmatrix} 1 & 0\\ \sin\theta & 1 \end{bmatrix} \begin{bmatrix} 1 & -\tan(\frac{\theta}{2}) \\ 0 & 1 \end{bmatrix}. \end{aligned}\]
[^4]: The curl of a tensor field \(\bm T\) is defined as \[\begin{aligned}
(\Curl T)\bm n := \Curl (\bm T^{\rm T} \bm n), \notag \end{aligned}\] where \(\bm n\) is an arbitrary constant vector, and the curl on the right-hand-side of the above equation is the curl of a vector field defined as \((\Curl v)_i = \epsilon_{ijk} v_{j,k}\), for any vector field \(\bm v\).
[^5]: The distribution of a random variable whose logarithm is distributed normally is called a log normal distribution. The cumulative distribution function of a log normal random variable with parameters \(\sigma\) and \(\mu\) is given by \[\begin{aligned}
\Phi \left ( \frac{\ln x-\mu}{\sigma} \right), \end{aligned}\] where \(\Phi\) is the cumulative distribution function of the standard normal distribution. | {'timestamp': '2017-06-07T02:04:19', 'yymm': '1706', 'arxiv_id': '1706.01646', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01646'} |
null | null |
# INTRODUCTION {#intro}
Coronal rain refers to cool and dense elongated plasma blobs or thread segments, which suddenly appear in the low corona, falling along coronal loops all the way down to the solar surface. This phenomenon was first recorded and classified as coronal sunspot prominences, commonly originating in coronal space and pouring down to spot regions, and was later observed and clarified as "coronal rain" by its characteristic feature of rapid brightening when approaching the chromosphere in the early 1970s. It has been well observed in spectral lines sampling temperature ranges from transition region (TR) to chromosphere . High-resolution spectroscopic observations found that coronal rain is composed of elongated blobs with average widths of 310 km, lengths of 710 km, average temperatures below 7000 K, average falling speed of 70 km s\(^{-1}\), and accelerations much smaller than the gravitational acceleration projected along loops. In non-flaring coronal loops, coronal rain is observed to appear during progressive cooling with high-speed downflow and dramatic evacuation in the loops . found spectroscopic evidence of plasma condensations taking place in coronal loops and forming coronal rain. found that coronal rain often occurs simultaneously in neighboring magnetic strands, forming groups of condensations which are seen as large clumps if they are close enough in proximity. These are called "showers" and can have widths up to a few Mm. found that coronal rain is a highly multi-thermal phenomenon with the multi-wavelength emission located very closely. The fragmentary and clumpy appearance of coronal rain blobs at coronal heights becomes more continuous and persistent at chromospheric heights immediately before impact. Rain clumps appear organized in parallel strands with density inhomogeneities in both chromospheric and TR temperatures. The electron density of rain clumps was found to be about 1.8-7.1\(\times10^{10}\) cm\(^{-3}\), through estimation based on absorption in multiple extreme ultraviolet (EUV) channels. Besides this quiescent coronal rain, which occurs in nonflaring coronal loops with relatively weak variation of energy and mass, flare-driven coronal rain, which appears in postflare loops as a result of catastrophic cooling, often emerges as a bunch of parallel strands extending from loop top to footpoint. With the phenomena of progressive cooling and in situ condensation, coronal rain is believed to be a representative of the general phenomenon of radiative thermal instability in an astrophysical plasma, that occurs whenever energy losses due to radiation overcome the heating input and the resulting cooling further amplifies energy losses. Numerical simulations firstly performed in one-dimensional setups [@Anti99; @Karpen01; @Muller03; @Muller04; @Muller05; @Xia11] have demonstrated that a heating input of a coronal loop concentrated at both footpoints, in chromosphere and near TR, can cause the loop to experience thermal non-equilibrium and thermal instability. This leads to coronal rain formation or prominence formation depending on the magnetic configuration of the loop. Footpoint heating of coronal loops is supported by observational evidence and by hydrostatic coronal loop models . Three-dimensional (3D) magnetohydrodynamics (MHD) numerical simulations found a concentration of Joule heating toward the upper chromosphere, TR, and lower corona as a result of the braiding of the magnetic field decreasing exponentially with height. Observations have shown a ubiquitous coronal mass supply where chromospheric plasma is accelerated and heated in type II spicules upward into the corona, and then leads to fading of type II spicules in chromospheric spectral line images . Inspired by the first two-dimensional (2D) evaporation-condensation model demonstrating prominence formation in a magnetic arcade, presented the first 2D coronal rain simulation, in which a large zigzag shape condensation forms across the top regions of a magnetic arcade covering a range of several Mm, then splits into many small blobs with elongated side wings and V-shaped features, descending along both sides of the arcade field lines. They performed statistic analyses to quantify blob widths and lengths, which average 400 km and 800 km, respectively, and velocities up to 65 km s\(^{-1}\), which is smaller than the observed average speed. Later, extended their model with an increased resolution of 20 km per cell and much longer time coverage, and studied the blob condensation in detail and found recurrent coronal rain showers to occur in limit cycles. They quantified the thermal structure of blob-corona TR and the variations of density, kinetic energy, and temperature during the impact of rain blobs on the chromosphere. They also showed how high-speed anti-parallel shear flows at two sides of condensations are induced that further facilitate fragmentation of the condensations. performed 3D simulations on coronal condensation in a magnetic configuration of a quadrupolar arcade system. In their models, coronal condensations occur in nearly horizontal coronal loop parts, where dense blobs first descend through (weak) local magnetic field undergoing Rayleigh--Taylor instability (RTI). Later on, small blobs start to follow more closely the magnetic field lines in the lower regions near the footpoints and slide down to the chromosphere, behaving like coronal rain. Their magnetic configuration did not represent a typical solar active region where coronal rain preferentially occurs. In a magnetic structure built from an observed magnetogram, simulated an active region validated with simulated EUV emissions. They found thermal non-equilibrium in a coronal loop leading to cooling and coronal condensation, although their spatial resolution was not enough to obtain coronal rain blobs. In order to better understand coronal rain in realistic coronal loops, we perform a 3D MHD simulation on quiescent coronal rain in a bipolar magnetic field, which is an idealized magnetic structure of a two-sunspot solar active region. The numerical methods and the simulation strategy are explained in Section [2](#sec:method){reference-type="ref" reference="sec:method"}. We present the results in Section [3](#sec:results){reference-type="ref" reference="sec:results"} and end with a conclusion and discussion in Section [4](#sec:conc){reference-type="ref" reference="sec:conc"}.
# Numerical method {#sec:method}
We setup a 3D Cartesian simulation box with extensions in \(x\) from-20 Mm to 20 Mm, in \(y\) from-30 Mm to 30 Mm, and in \(z\) from 0 Mm to 60 Mm. The solar gravity points along the negative \(z\) direction and the \(z=0\) bottom surface represents the solar surface. With five levels of adaptive mesh refinement, the mesh has an effective resolution of \(384\times576\times768\), with smallest grid cell size of 78 km. To mimic a magnetic configuration of a simple two-sunspot active region, we use a potential magnetic field composed of two dipoles with vertical and oppositely directed moments located below the solar surface at (0, \(y_c\), \(-z_c\)) and (0, \(-y_c\), \(-z_c\)), respectively: \[\begin{aligned}
\begin{aligned} & B_x=B_{x+}+B_{x-},~~B_y=B_{y+}+B_{y-},~~B_z=B_{z+}+B_{z-},\\ & B_{x\pm}=3x(x+z_c)f_{\pm},\\ & B_{y\pm}=3(y \mp y_c)(z+z_c)f_{\pm},\\ & B_{z\pm}=[2(z+z_c)^2-x^2-(y\mp y_c)^2]f_{\pm},\\ & f_{\pm}=\frac{\pm B_c}{[x^2+(y\mp y_c)^2+(z+z_c)^2]^{2.5}}, \end{aligned} \end{aligned}\] where \(y_c=10\) Mm, \(z_c=15\) Mm, and \(B_c=200\) G. This magnetic setup gives maximal 120 G on the bottom surface (\(z=0\)) and about 3 G at a height of 30 Mm. Although the magnetic field strength is much weaker than a typical solar active region, these weak magnetic loops may be found in a decayed active region. To construct a solar atmosphere, we let the model obtain a realistic thermal and gravitational stratification by relaxing an initial state, which has a vertical distribution of temperature of about 8000 K below the TR at a height of 2 Mm with smooth connection (TR at 160,000 K using hyperbolic tangent function) to the upper region where the temperature increases with height in such a way that the vertical thermal conduction flux is always \(2\times10^5\) erg cm\(^{-2}\) s\(^{-1}\). The density is then calculated solving a hydrostatic equation with the number density at the bottom being \(7.3\times10^{12}\) cm\(^{-3}\). Our simulations are performed by solving the thermodynamic MHD equations given by \[\begin{aligned}
\frac{\partial \rho}{\partial t}+\nabla\cdot\left(\rho\mathbf{v}\right)&=0,\\ \frac{\partial \left(\rho\mathbf{v}\right)}{\partial t}+\nabla\cdot\left( \rho\mathbf{vv}+p_{\rm tot}\mathbf{I}-\frac{\mathbf{BB}}{\mu_0}\right)&=\rho\mathbf{g},\\ \frac{\partial E}{\partial t}+\nabla\cdot\left(E\mathbf{v}+p_{\rm tot}\mathbf{v}- \frac{\mathbf{BB}}{\mu_0}\cdot\mathbf{v}\right)&=\rho\mathbf{g}\cdot \mathbf{v}+H-R+\nabla\cdot\left(\boldsymbol{\kappa}\cdot\nabla T\right), \\ \frac{\partial \mathbf{B}}{\partial t}+\nabla\cdot\left(\mathbf{vB}-\mathbf{Bv}\right)&=0, \end{aligned}\] where \(\rho\), \(\mathbf{v}\), \(\mathbf{B}\), and \(\mathbf{I}\) are the plasma density, velocity, magnetic field, and unit tensor, respectively. Moreover, \(p_{\rm tot}\equiv p+B^2/2\mu_0\) is total pressure with gas pressure \(p=2.3 n_{\rm H} k_{\rm B} T\) assuming full ionization and an approximate helium abundance (\(n_{\rm He}/n_{\rm H}=0.1\)), \(E=p/(\gamma-1)+\rho v^2/2+B^2/2\mu_0\) is total energy, and \(\mathbf{g}=-g_\odot r_\odot^2/(r_\odot+z)^2\mathbf{\hat{z}}\) is the gravitational acceleration with solar radius \(r_\odot\) and the solar surface gravitational acceleration \(g_\odot\). We use normalization units of length \(L_0=10\) Mm, time \(t_0=1.43\) min, temperature \(T_0=10^6\) K, number density \(n_0=10^9\) cm\(^{-3}\), velocity \(v_0=116.45\) km s\(^{-1}\), and magnetic field \(B_0=2\) G to normalize the equations. Any number without a specified unit has the according normalization unit thoughout the paper. We use the Adaptive Mesh Refinement Versatile Advection Code (MPI-AMRVAC) to numerically solve these equations with a scheme setup combining the Harten--Lax--van Leer Riemann solver with a third-order slope limited reconstruction and a three-step Runge--Kutta time integration. We add a diffusive term into the induction equation to keep the divergence of magnetic field under control. Thermal source terms in the energy equation, such as field-aligned (\(\boldsymbol{\kappa}=\kappa_\parallel \mathbf{bb}\)) thermal conduction and optically thin radiative cooling \(R=1.2 n_{\rm H}^2\Lambda(T)\), are treated as in our previous studies using a Super Time Stepping scheme and an exact integration scheme, respectively. To establish a hot corona, the global coronal heating is simulated by adding a parametrized heating term \(H_1=c_1 B^{1.75} n_{\rm e}^{0.125}/r^{0.75}\) erg cm\(^{-3}\) s\(^{-1}\), where \(c_1=4.54\) is a constant, \(B\) is magnetic field strength, \(n_{\rm e}=1.2 n_{\rm H}\) is number density of electrons, and \(r=1/|(\nabla \mathbf{b}) \cdot \mathbf{b}|\) is the local radius of curvature of magnetic field lines, where \(\mathbf{b}=\mathbf{B}/B\) is a unit vector along the magnetic field. We split off the potential magnetic field as a time-invariant component, and numerically solve for the deviation from it to avoid negative pressure problems caused by errors in magnetic energy when magnetic energy dominates the internal energy in extremely low-\(\beta\) plasma. On the four side boundaries, we use a symmetry boundary condition for density and gas pressure, and extrapolate magnetic field according to a zero-gradient assumption. On the bottom boundary, we fix the density, pressure, and magnetic field the same as in the initial condition. On the top boundary, we solve the hydrostatic equation to obtain density and pressure values assuming a continuous temperature distribution, and adopt the same treatment for magnetic field as on side boundaries. We use anti-parallel symmetry conditions for velocity to ensure zero velocity on the all boundary surfaces. The normal component of extrapolated magnetic field is modified to numerically fulfill a centered difference zero-divergence constraint. The initial setup is not in an equilibrium, so we first run the simulation with only global heating in a relaxation period of 114 min to reach a quasi-equilibrium. Starting from that state, we reset the time to zero and add a localized heating to evaporate chromospheric plasma into the coronal loop to cool down to coronal rain. The localized heating \(H_2\) is concentrated at both foot point chromospheric regions of the arcade with strong vertical magnetic field and decays with height above \(z=4\) Mm in a Gaussian profile: \[H_2=\begin{cases}f(t) c_2 (B_z/B_h)^2 e^{-((z-z_h)/H_m)^2} & \text{if \(z > z_h\)} \\ f(t) c_2 (B_z/B_h)^2 & \text{if \(z \le z_h\)}, \end{cases}\] where \(c_2=10^{-2}\) erg cm\(^{-3}\) s\(^{-1}\), \(B_h=68\) G, \(z_h=4\) Mm, \(H_m=3.16\) Mm, and \(f(t)\) is a linear ramp function of time to switch on the heating smoothly and keep it fixed afterwards. This heating mimics long-term mild heating in nonflaring coronal loops, and can be modified to represent impulsive violent heating of flares to stimulate postflare loops and flare-driven coronal rain in the future. This final stage is simulated for 114 min, and we present the results in the following sections.
# RESULTS {#sec:results}
The localized heating evaporates chromospheric plasma into hot coronal plasma and causes upflows from footpoint regions with speeds peaking at about 50 km s\(^{-1}\) in the lower corona. A loop top region centered at (0,0,40) Mm takes the lead in cooling because the radiative cooling there dominates over heating and thermal conduction as a result of density enhancement. This enhancement is caused by the evaporated flows along the heating loops and by squeezing due to the expansion of underlying heated loops. At 13.5 min, the central region cools down to 0.02 MK, below which temperature our radiative cooling vanishes. Then, induced siphon flows, with speeds up to 24 km s\(^{-1}\) coming from two sides, merge there and create a high-density low-temperature plasma blob. This blob grows in three dimensions and the most rapid extension happens along the direction parallel to the polarity inversion line (\(x\)-axis) as shown in Figure [\[fig:first\]](#fig:first){reference-type="ref" reference="fig:first"}. This is caused by the similarity and concurrency of the condensation process across the loops with a similar shape and height distributed along the \(x\)-direction. Observations also found that condensations in neighboring magnetic field strands occur simultaneously, which was thought to be due to a common footpoint heating process. The head-on impact of siphon flows not only compresses the central cool plasma to form a dense plasma blob, but also at the same time launches two rebounding slow shocks propagating against the continuous siphon flows along the magnetic field. The shock fronts are clearly seen in the temperature and density maps on the slices through the condensation as shown in Figure [\[fig:first\]](#fig:first){reference-type="ref" reference="fig:first"} (b)--(d). The shock-front surfaces are curved, slow shock fronts and show up on a bundle of similar magnetic flux tubes where condensations occur successively. The processes of condensation and shock formation on these flux tubes are essentially the same, although they start at different times. Therefore, the whole history of the condensation process in a single flux tube is represented by snapshots like Figure [\[fig:first\]](#fig:first){reference-type="ref" reference="fig:first"} (b)--(d). Similar rebound shocks were also found in our previous 2.5D simulations. At this moment, the shocks decrease the inflow speed from about 50 km s\(^{-1}\) to about 17 km s\(^{-1}\), compress the plasma from \(5\times10^{8}\) to \(9\times10^{8}\) cm\(^{-3}\), and heat the plasma from 0.18 MK to 0.21 MK. As the condensation extends with a span of 37 Mm across loop tops, it starts to distort and becomes fragmented in about 6 min due to RTI, which is shown in Figure [\[fig:RTI\]](#fig:RTI){reference-type="ref" reference="fig:RTI"} where the vertical velocity pattern alternates upflows with downflows in a direction perpendicular to the magnetic field. The dense coronal rain plasma is gathered in falling spikes and upflows correspond to hot coronal plasma. The plasma \(\beta\) of the condensation is about 0.5 and the magnetic field strength is about 2 G. The Atwood number \(\mathbf{A}\), which is \((\rho_{high}-\rho_{low})/(\rho_{high}+\rho_{low}),\) indicating the density disparity between two layers of fluid under acceleration in classical RTI theory, for the coronal rain plasma and underlying coronal plasma is about 0.93. The coronal rain plasma is about 25 times denser than the underlying coronal plasma. Similar results were found in our previous studies. As shown in Figure [\[fig:RTI\]](#fig:RTI){reference-type="ref" reference="fig:RTI"} (c), the falling spikes of coronal rain plasma have larger growth rates and penetrate deeper into the opposite region than rising bubbles of coronal plasma as expected for Atwood numbers close to 1. We note that the RTI modes in this study are almost perpendicular to local magnetic field, which should locally behave as in hydrodynamics where shortest wavelengths grow the fastest. However, since the magnetic field lines are line-tied at the bottom, the interchange deformation of magnetic field lines and the RTI mode growth is restrained. The fragmented blobs move downward in three groups, namely, two large clumps at the two flanks where coronal loops are inclined, and one near the central vertical loops, as shown in the first row of Figure [\[fig:evolve\]](#fig:evolve){reference-type="ref" reference="fig:evolve"}. These blobs initially fall roughly in the vertical plane (\(y=0\)) along the PIL. Because the evolution of our bipolar system is roughly symmetric and magnetic field is weak enough at those loop top regions, the blobs, sitting at loop tops with nearly zero projected gravity along the loops, can displace magnetic field transversally dragging down their hosting magnetic loops. Blobs of the central group fall slower than those at flanks due to the stronger resistant magnetic-tension force they encounter. They fragment further into smaller ones. Many tiny blobs gradually evaporate to hot plasma and disappear in the density isosurface view. The flank clumps fall almost vertically with speeds about 16 km s\(^{-1}\) until they reach a lower height of about 13 Mm. Then, they start to be more constrained in their motions by the local magnetic field with plasma \(\beta\) about 0.1 and to be forced to stream down along the magnetic loops toward the foot points in the negative polarity region. After the flank blobs drain down completely, coronal rain blobs mainly form at the top region of the central loops and fall along the loops intermittently (see the second row of Figure [\[fig:evolve\]](#fig:evolve){reference-type="ref" reference="fig:evolve"}). During their fall, the condensed blobs keep changing shape, elongating, separating, or increasing in size. In order to understand the time evolution, we quantify several properties of the coronal rain blobs in terms of mass, mass drainage rate, mean plasma \(\beta\), angle between local velocity and magnetic field vector, instantaneous mean speed and maximal speed, and plot the time evolution curves in Figure [\[fig:tcurve\]](#fig:tcurve){reference-type="ref" reference="fig:tcurve"}. Coronal rain blobs are cool and dense plasma above transition region, so we practically define that they are composed of cells with density larger than \(7\times10^{10}\) cm\(^{-3}\), with temperature lower than 0.1 MK, and with height larger than 6 Mm. The mass drainage rate is the integrated coronal rain mass flux through a horizontal plane at 6.5 Mm height. During the first condensation from 14 min to 30 min, the condensation has a roughly linear growth rate of \(2.84\times10^{9}\) g s\(^{-1}\), which is much slower than the one in 3D prominence simulation (\(1.14\times10^{10}\) g s\(^{-1}\)) and the estimation from prominence observations (\(1.2\times10^{10}\) g s\(^{-1}\)) . Comparing with our previous 2D coronal rain simulations (assuming 1000 km integral depth in the third direction), the condensation rate in this 3D model is an order of magnitude faster. From 30 min to 38 min, the growth rate slows down because the fully developed RTI breaks the initial large blobs into pieces in which tiny blobs and very thin threads are heated to hotter and less dense material than coronal rain plasma mainly by thermal conduction. From 38 min until the first peak of the mass curve at 47.8 min, the growth rate is about \(3.25\times10^{9}\) g s\(^{-1}\), slightly faster than in the initial phase. After the first peak, the mass begins to drop rapidly when fast drainage due to falling of coronal rain starts. This fast drainage, which finishes at 60 min, is contributed by the large falling clumps at the two flanks. After that, smaller coronal rain blobs formed at the central loop parts stream down episodically. As continuous condensation increases the coronal rain mass, the episodic fallen blobs, which correspond to the peaks in the mass drainage curve, decrease the mass. Therefore, the total mass remains at a steady level for a while from 55 min to 85 min. After that, the mass increases and decreases forming a large hill in the curve over a period of 25 min when a rain shower forms and falls. The plasma \(\beta\) of the condensation is relatively high (more than 0.5) at the beginning as the initial condensation forms in a weak magnetic field region at the loop top. With the solar gravity, RTI is inevitable under these circumstances and the nearly vertical falling of the condensation spikes pushes the magnetic loops down, making the angle between velocity and magnetic field close to 80 degrees. As the coronal rain blobs fall to lower regions with stronger magnetic field and smaller plasma \(\beta\), they are ultimately forced to follow the guide of magnetic field lines with the angle decreasing. Therefore, the curve of plasma \(\beta\) and the angle match each other as shown in Figure [\[fig:tcurve\]](#fig:tcurve){reference-type="ref" reference="fig:tcurve"} (b). The mean speed of coronal rain plasma at each instant shifts from about 20 km s\(^{-1}\) to about 30 km s\(^{-1}\) after the fall of the large flank blobs, while the instantaneous maximal speed, which is about two to three times larger than the instantaneous mean speed, also increases to about 100 km s\(^{-1}\). We wrote a program to automatically count the coronal rain blobs in each snapshot and quantify their properties, such as number, mass, volume, density, centroid location, and centroid speed. The key algorithm is to group adjacent cells that contain coronal rain plasma into one blob, above TR, which is detected based on the vertical gradient of local density and temperature. To consider fully resolved blobs, we neglect tiny blobs containing less than 64 cells at the highest refinement level. Using this program, we performed statistical analysis of coronal rain blobs as shown in Figure [\[fig:statistic\]](#fig:statistic){reference-type="ref" reference="fig:statistic"}. The total number of blobs in all 352 snapshots is 2406. The mean number of blobs per snapshot is 6.8. The maximal number of blobs in one snapshot is 17. In the curve of blob number, we can see that the initial big blob breaks into three pieces at 30 min. Then the number increases up to 17 at 55 min and decreases in the later phase. The total mass of coronal rain blobs matches well with the mass curve in Figure [\[fig:tcurve\]](#fig:tcurve){reference-type="ref" reference="fig:tcurve"} (a). Blobs are dominated by small ones with mass less than \(5\times10^{10}\) g with the heaviest blob of \(3.2\times10^{12}\) g. The distribution of speed has a peak around 15 km s\(^{-1}\) and a long tail over 100 km s\(^{-1}\). The mean speed of all blobs is 22.4 km s\(^{-1}\). The blob with the maximal speed of 104.5 km s\(^{-1}\) is very small with a mass of \(2.4\times10^{9}\) g. After analyzing the overall properties of the modeled coronal rain event, we go into more detail with case studies of typical coronal rain blobs. In Figure [\[fig:fallblob\]](#fig:fallblob){reference-type="ref" reference="fig:fallblob"}, we find several coronal rain blobs, presented by red density isosurfaces, at different heights in different time stages of their life. We focus on a stretched coronal rain blob in the lower left part of panel (a). It is falling roughly in the central plane with \(x\) close to 0. It has a swelled head with relatively high speed (50 km s\(^{-1}\)) before hitting the TR, a straight elongated body, and a hook-like tail. The blob is falling along a bundle of magnetic loops with its lower parts residing on the shorter loops, as shown by two field lines going through the head and the tail part. A vertical slice cutting through the blobs shows the \(v_y\) distribution. There are fast following flows in the wake of the falling blob. The elongating blob has low gas pressure in its body inducing siphon flows that enhance the wake flows (see the blue regions). The siphon flows upstream of the falling blob become neutralized by the pushing motion of the blob. Blobs near the loop top in higher regions have little motion along the loops, so the siphon flows from both sides towards the blobs are obvious and bring material to feed the growth of the blobs. Once the blobs reach the low-\(\beta\) region, the lower part at the lower loops feels stronger projected gravity due to the shape of the loops. The lower part falls faster than the upper part along the loops, so an initially compact blob gets elongated and this tearing also decreases the gas pressure in the straight body of an elongated blob. Besides, when a plasma blob streams down in a flux tube with narrower cross-section closer to the footpoints, it will be squeezed and elongated. We made a small horizontal slice cutting across the selected blob, and plot the temperature map (panel (b)) and \(v_z\) map (panel (c)) with number density contours on them. The blob is wrapped by a TR which connects the 0.04 MK coronal rain plasma to 1 MK coronal plasma in a range of about 500 km. The thickness of the blob in this cross-section is about 700 km if we count the 7 \(n_0\) density contour and 0.1 MK temperature contour as its border. The density contours and the temperature map show increasing density and decreasing temperature when getting closer to the center of the blob, which reflects the multi-density and multi-thermal nature of the coronal rain blobs. In the vertical velocity map, the downflow region occupies not only the coronal periphery but also half the blob itself. These anti-parallel shearing flows are divided over the core of the blob, which leads to elongation of the blob during its fall. The large velocity difference between different parts of individual blobs is also reflected in the significant difference between the maximal speed and the mean speed at each snapshot in Figure [\[fig:tcurve\]](#fig:tcurve){reference-type="ref" reference="fig:tcurve"} (c). To directly compare with observed images from the Solar Dynamics Observatory Atmospheric Imaging Assembly (AIA) instrument, we made synthetic observations on our simulations in selected viewing angle and retrieved synthetic images with a technique described in detail in . To mimic absorption of EUV emission by the core plasma of condensations, we exclude emission coming from behind dense plasma with a number density larger than \(2\times10^{10}\) cm\(^{-3}\), except for the 304 Å wave channel. For example, Figure [\[fig:fsyn\]](#fig:fsyn){reference-type="ref" reference="fig:fsyn"} shows synthetic views on the modeled coronal rain and coronal loops at 60.4 min in 211 Å, 193 Å, 171 Å, and 304 Å wave channels, which have main contribution temperatures of about 1.8, 1.5, 0.8, and 0.08 MK, respectively. The line of sight of Figure [\[fig:fsyn\]](#fig:fsyn){reference-type="ref" reference="fig:fsyn"} has an azimuthal angle of 45\(^{\circ}\) away from the PIL and an elevation angle of 20\(^{\circ}\). In hot channels panel (a), (b), and (c), the bright coronal loops and their footpoint regions appear more dispersed in higher-temperature channels. Dark clumps of coronal rain reside in dim loops in 211 Å, while adhering to bright halo segments of 171 Å loops. Bright clumps of coronal rain are found in the cooler 304 Å channel sampling TR plasma. The impact site of coronal rain at the chromosphere is at this moment a fairly compact region of negative polarity with a diameter less than 10 Mm. In the movie of Figure [\[fig:fsyn\]](#fig:fsyn){reference-type="ref" reference="fig:fsyn"}, the inital condensation is barely seen in 304 Å , because the density of the emission-contributing plasma (around 0.08 MK) is relatively low (about \(2\times10^9\) cm\(^{-3}\)) at the transition layer between the condensation and corona. In the last phase of the simulation from 90 min to 113 min, a bursty condensation makes a coronal rain "shower" in a large bundle of coronal loops. The total mass of the condensation increases and peaks at 105 min. The total number of blobs first increases and soon decreases because neighboring small condensations grow fast enough to touch one another and to be counted as one blob. We plot a representative snapshot of this phase in Figure [\[fig:fshower\]](#fig:fshower){reference-type="ref" reference="fig:fshower"}. As shown by density isosurfaces in panel (a), long streams of condensed plasma are connected to one another. The horizontal extension of the streams along the \(x\)-direction, covering loops with similar length, is about 10 Mm, while the thickness of the streams along \(y\)-direction is around 1 Mm. In panel (b), a horizontal cross-section of the streams finds the density inhomogeneities within the coronal rain clumps, which is similar to the results found in observations, although the spatial sizes are different. In our model, a coronal rain clump can have density inhomogeneity because it is composed of several seed blobs which grow and move along the magnetic loops closely in space and time. In panel (c), a synthetic view in AIA 171 Å channel shows dark clumps, which are the high-density (\(2\times10^{10}\) cm\(^{-3}\)) cores of coronal rain streams in panel (a), in bright coronal loops.
# Conclusions and discussion {#sec:conc}
We realized a 3D coronal rain simulation with unprecedented high resolution and a more realistic magnetic environment than previous models. The formation process of coronal rain, characterized by progressive cooling and condensation in situ, is reproduced in our model. The physical reason behind it is the radiation-dominated thermal instability. The formation of the first condensation is presented in detail with its 3D shape extending along the PIL and with the shape of rebounding shock fronts indicating the time sequence of the condensation progress. The first long condensation on the loop top region develops RTI interchange deformation and becomes fragmented into smaller blobs. The blobs fall slowly and vertically against magnetic field loops until reaching low-\(\beta\) regions and start to slip and fall along the field loops from loop top to loop footpoints. A similar phenomenon was presented by in a 3D MHD simulation of the falling of dense fragments along magnetic flux tubes of an erupted solar filament, in which the initial falling speed of the dense blobs is about 300 km s\(^{-1}\), much higher than our case, and fast enough to generate slow shocks, which are not found ahead of the falling blobs in our study. They also found that these fast-moving blobs with imposed initial motions misaligned with magnetic field soon become deformed and mixed due to the feedback from dragged magnetic field lines . Similarly, the misaligned motion of our falling blobs near the loop tops also leads to reshaping and fragmentation of the blobs. Statistical study of the coronal rain blobs found in all snapshots in our simulation shows that small blobs with mass less than \(10^{10}\) g dominate the population despite the fact that unusually large blobs show up in the initial and last phases. Large blobs have statistically slower centroid speed than small ones. When the blobs stream along the magnetic loops, they are stretched and develop a non-uniform velocity pattern. The elongated blobs have faster (siphon) flow on those parts closer to the center of curvature of the hosting coronal loops. AIA synthetic observations of our simulated coronal rain well resemble real observations in several aspects, like dark falling clumps in hot EUV channels and bright rain blobs in cool 304 Å channel. We also find density inhomogeneities within the coronal rain clumps in a rain "shower", which may explain the observed multi-stranded nature of coronal rain. The magnetic field strength in this model is as yet relatively low compared to typical active region coronal loops at the same height. It is computationally challenging to simulate plasma with extremely low \(\beta\) and high Alfvén speed when solving conservative energy equations and using explicit schemes. At the cost of much more computational resources, we can get closer to the realistic magnetic field strength in the work in progress. The RTI interchange mode of loop-top condensations may then be suppressed by the rigid line-tied coronal loops in very low-\(\beta\) regimes. The physics of rain-blob formation and the dynamics of coronal rain blobs in the lower regions should remain the same. Since the RTI interchange mode of prominence plasma has been observed and modeled in the quiescent region, it may be found on high-\(\beta\) loop-top by future observations. Previous simulations of RTI in prominences start with existing static prominence material excluding the dynamic formation process; the model here is the first one covering both formation and subsequent RTI of prominence-like plasma in similar local magnetic environment, for example, weak and locally horizontal magnetic field. However, the vertical thickness of the initial condensation is only about 2 Mm, much smaller than a typical prominence and the bipolar magnetic structure is quite different from prominence-hosting magnetic structures with magnetic dips, thus the details of RTI here are different from those in modeled prominences and observed prominences. We further plan to use realistic magnetic field extrapolated from actual observed magnetograms and study the dynamics of coronal rain blobs. Flare-driven coronal rain has not yet been simulated and needs simulations for further understanding of its fine-scale strands and the cooling of post-flare loops. | {'timestamp': '2017-06-07T02:07:25', 'yymm': '1706', 'arxiv_id': '1706.01804', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01804'} |
null | null |
null | null |
# Introduction {#sec:intro}
Adiabatic quantum computation (AQC) has been proposed as a successful technique for solving certain classes of optimization problems (see ). In particular, quantum annealers (QA) such as the ones manufactured by D-Wave Systems Inc. approximate the adiabatic evolution in the presence of various sources of noise and finite temperature to solve Ising models of the form \[\label{Ising_def} \tag{Ising} \min_{s \in \mathbb \{\pm 1\}^n}\, s^t J S + h^t s\,,\] where diagonal entries of \(J\) are zeroes, and the nonzero entries of \(J\) create a subgraph of a sparse graph called a *Chimera* graph. Many researchers have explored the applicability of quantum annealing to more-general optimization problems. Solving Ising models without the Chimera graph structure on a QA is discussed in. Degree reduction techniques, such as the one described in, enable us to solve unconstrained binary polynomial programming problems of higher order using a QA. Several studies have employed quantum annealing for solving real-world applications, and discusses the possibility of using QAs for solving constrained problems. Many optimization problems, on the other hand, involve integer-valued variables beyond binary values. Examples of such problems could be the number of vehicles or products travelling along each route of a supply chain (see ), or the number of traded assets in portfolio optimization (see ). This paper focuses on solving quadratic problems on bounded integer domains, alternatively referred to as unconstrained integer quadratic programming (UIQP) problems, i.e., problems of the following form: \[\label{UIQP} \tag{UIQP} \begin{array}{lll} \min & x^T Q x + q^t x,\\ \text{s.t.} &x_i \in \{0, 1, 2, \ldots, \kappa^{x_i} \} & \quad \text{for } i=\{1,2, \ldots, n\}, \nonumber \\ \end{array}\] where \(\kappa^{x_i} \in \mathbb Z_+\) is an upper bound on \(x_i\); here, \(\mathbb Z_+\) denotes the set of non-negative integers. Note that in problems where \(x_i \in \{ \alpha, \alpha+1, \alpha+2, \ldots, \alpha+\kappa^{x_i}\}\), we may shift \(x_i\) and substitute it with \(x_i-\alpha\); hence, without loss of generality, we may assume that the domain of \(x_i\) starts at 0. It is worth mentioning that, although current QAs are limited to representing Ising models, an alternative approach to encoding bounded integer variables in terms of spin variables is to consider physical implementations of quantum processors as representing Potts-Ising models. Our approach is different and is based on reformulating a UIQP problem as an Ising model. To this end, we represent each integer variable as a linear combination of several binary variables, i.e., \[x_i = \sum_{j=1}^{d^{x_i}} c^{x_i}_j y^{x_i}_j = (c^{x_i})^t y^{x_i}, \label{intEncoding}\] where \(c_j^{x_i} \in \mathbb Z_+\), \(y_j^{x_i} \in \{ 0, 1\}\) for \(j \in \{1, \ldots, d^{x_i}\}\), and the superscript \(x_i\) is used for clarity to denote that \(c^{x_i}\) and \(y^{x_i}\) correspond to the encoding of variable \(x_i\). This representation is referred to as *integer encoding*. Some of the well-known integer encodings are binary and unary encodings in which \(c_j^{x_i} = 2^{j-1}\) and \(c_j^{x_i}=1\), respectively. The *width* of an integer encoding, denoted by \(d^{x_i}\) in [\[intEncoding\]](#intEncoding){reference-type="eqref" reference="intEncoding"}, refers to the number of binary variables required for encoding integer variable \(x_i\). The width of binary and unary encodings are \(\left \lfloor \log_2(\kappa^{x_i}) \right \rfloor+1\) and \(\kappa^{x_i}\), respectively. Since a binary variable \(y_j\) could be represented with a spin variable \(s_j\) via the affine transformation \[\label{binToSpin} y_j = \frac{1}{2} \l ( s_j + 1\r),\] integer variable \(x_i\) could be encoded into several spin variables as \[\label{int2SpinEncoding} x_i = \frac{1}{2} \l ( \sum_{j=1}^{d^{x_i}} c_j^{x_i} + \sum_{j=1}^{d^{x_i}} c_j^{x_i} s_j^{x_i} \r),\] which enables us to successfully represent a UIQP as an Ising model. After derivation of an Ising model equivalent to [\[UIQP\]](#UIQP){reference-type="eqref" reference="UIQP"}, heuristic methods including quantum annealing could be employed to find the ground state of the problem. The performance of the heuristic methods highly depends on the energy landscape of the problem. In particular, landscapes with tall barriers are challenging for the hill-climbing heuristics, and having wide barriers could impact the performance of methods like quantum annealing that could potentially benefit from quantum tunnelling (see and references therein for more details). Having coefficients that are different in orders of magnitude is undesirable, as they could create energy landscapes with tall and wide barriers. Moreover, D-Wave quantum annealers have only a low precision of approximately \(10^{-2}\) for couplings' strengths and local fields' biases (i.e., entries of \(J\) and \(h\) in [\[Ising_def\]](#Ising_def){reference-type="eqref" reference="Ising_def"}), scaling \(J\) and \(h\) to a very limited range, (e.g., \([-2,2]\)). Several sources of noise such as thermal excitations and control errors contribute to the low precision of the machines; in, Albash et al. propose a noise model for D-Wave devices that includes the control noise of the local field and couplings of the chip. Zhu et al. show that increasing the classical energy gap beyond the intrinsic noise level of the machine can improve the success of the D-Wave Two QA. In, *resilience* is defined as a metric for measuring the resistance of a problem to noise. More precisely, resilience is the probability that the ground state does not change under random field fluctuations found on the chip. In this paper, we argue that restricting the range of the coefficients of a problem could improve resilience; and our results, presented in Section [4](#sec:res){reference-type="ref" reference="sec:res"} of this paper, support this argument Studies in suggest that QAs could be biased in finding degenerate ground states; in other words, sampling with QAs does not return different degenerate solutions with uniform probability. Moreover, degenerate ground states are easier to reach; therefore, benchmarks created for QAs tend to avoid degeneracy and have a unique ground state. While there is a need to rigorously study whether highly degenerate low-energy excited states could impact reaching the ground state, avoiding degeneracy could be an alternative. Recall that in integer encoding, an integer variable, \(x\), is substituted with \(c^t y\), where we assume \(c \in \mathbb Z_+^d\) and \(y \in \mathbb B^d\). Any binary vector \(y\) returns an integer value. If the total number of binary combinations, i.e., \(2^d\), is larger than the summation of entries of \(c\), i.e., \(\sum_{i=1}^d c_i\), some of the integers occur at more than one binary combination. The *code-word* of an integer value \(\chi\), for \(\chi \in \{1, 2, \ldots, \kappa\}\), refers to the cardinality of the set \(\{ y: c^t y = \chi \}\). Code-words more than one is what we refer to as *redundancy* in this paper. Under the assumption that the QA is inclined to observe integers with higher code-words, it is ideal in an integer encoding to have a unique code-word for all integers. Binary encoding, for example, keeps the code-word of each integer uniformly at one, but the code-word in unary encoding is highly variant and the redundancy has its peak at \(\l \lceil \frac{\kappa}{2} \r \rceil\) with word count \(\kappa \choose \lceil \frac{\kappa}{2}\rceil\). This work presents an integer encoding with minimal width that creates noise-resilient Ising models. Since one or several qubits is assigned to each spin variable and the number of qubits on a chip is limited, it is desirable to keep the size of the Ising model, and thus the width of the encoding, as small as possible. Binary encoding has the minimum width and uniform code-word; however, the coefficients \(c_j\)'s in the binary encoding could get arbitrarily large and result in an error-prone (noisy, non-resilient) Ising model. On the other hand, unary encoding does not expand the range of the coefficients of the problem, but it has an exceedingly large degree and redundancy. The notation used in this paper is as follows. Generally, we reserve the upper-case letters, lower-case letters, and Greek alphabet for matrices, vectors, and scalars, respectively, with only a few exceptions; these exceptions, however, should be clear from the context. An all-ones vector and the identity matrix are denoted by \(e\) and \(I\), respectively. Entries of a vector or matrix are differentiated with subscript indices. \(\lceil \cdot \rceil\) and \(\lfloor \cdot \rfloor\) denote the ceiling and floor of a number, respectively. The \(\log(\cdot)\) function is in base two; and \(\text{sgn}(\alpha)\) returns the sign of \(\alpha\). The function \(\delta(A): \mathbb R^{n\times n} \rightarrow \mathbb R^n\) returns a vector with the diagonal entries of \(A\); and the function \(\Delta(a): \mathbb R^n \rightarrow \mathbb R^{n\times n}\) returns a diagonal matrix with entries of \(a\) on the diagonal. This paper is organized as follows. In the next section, we present bounded-coefficient encoding under the assumption that an upper bound on the coefficients of the integer encoding is available. In Section [3](#sec:findingUB){reference-type="ref" reference="sec:findingUB"}, we present how, given the precision, we may find an upper bound on the coefficients of the integer encoding. We test our algorithm by comparing binary and bounded-coefficient encodings in Section [4](#sec:res){reference-type="ref" reference="sec:res"}. Finally, we conclude our discussion in Section [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}.
# Bounded-Coefficient Encoding {#sec:bounded-coeff encoding}
Let \(x\) be an integer variable with upper bound \(\kappa^{x}\), and \(\mu^{x}\) be an upper bound on the coefficients of the integer encoding. In other words, we represent \(x\) as \[\label{intEncoding_sec2} x = c^t y,\] where \(y\in \mathbb B^{d^{x}}\), \(c \in \mathbb Z_+^{d^x}\), and \(c_i \le \mu^x\) for \(i =1, \ldots, d^x\). The encoding we propose in this paper, which we refer to as *bounded-coefficient encoding*, is summarized in the following algorithm:
\[\label{bounded-coeff_case0} c^x= \l [2^0, 2^1, \ldots, 2^{\lfloor \log (\kappa^x) \rfloor-1}, \kappa^x-\sum_{i=1}^{ \lfloor \log (\kappa^x) \rfloor} 2^{i-1} \r ]\]
\[\label{bounded-coeff_eq} c^x_i=\begin{cases} 2^{i-1} & \text{ for } i=1, \ldots, \rho\\ \mu^x & \text{ for } i=\rho+1, \ldots, \rho + \eta \\ \nu-\eta \mu^x & \text{ for } i=\rho + \eta +1 \text{ if } \nu-\eta \mu^x \neq 0 \end{cases}\]\
By the first part of our encoding, i.e., \(1, \ldots, 2^{\rho-1}\), we can encode all integers \(\{ 0, 1, \ldots, \mu^x\}\). By adding \(\mu^x\) to those binary combinations, we can generate all integers from \(\mu^x\) to \(2\mu^x\); if we add two \(\mu^x\) factors, we get integers \(2\mu^x\) to \(3\mu^x\), and so on. Finally, we are guaranteed that we get integers \(\eta\mu^x\) to \(\kappa^x\) because \(\kappa^x-\eta\mu^x < \sum_{i=1}^{\rho} 2^{i-1}-\mu^x\) by the fact that \(\nu-\eta \mu^x < \mu^x\).\
As mentioned earlier, a potentially advantageous property of an integer encoding is uniform or low-variant redundancy. In general, the bounded-coefficient encoding has redundancy because we have more than one coefficient with a value of \(\mu\). It is, however, worth mentioning that forcing the following constraints on the binary variables of [\[intEncoding_sec2\]](#intEncoding_sec2){reference-type="eqref" reference="intEncoding_sec2"} makes the word count of each integer value unique: \[\begin{array}{rcll} \sum_{j=1}^{\rho} c_j y_j & \ge & (2^{\rho}-\mu^x) y_i & \text{for } i=\rho +1, \\ y_{i} &\ge& y_{i+1} & \text{for } i=\rho+1, \ldots, d^x-1,\\ \sum_{j=1}^{\rho} c_j y_j &\ge& (2^{\rho}-c_{d^x}) y_{d^x}. \end{array}\] Note that \((2^{\rho}-\mu^x) \le \mu^x\); however, \((2^{\rho}-c_{d^x})\) may not necessarily be less than or equal to \(\mu^x\). If we wish to limit the coefficients of our constraints to \(\mu^x\), we can substitute \((2^{\rho}-c_{d^x}) y_{d^x}\) in the right-hand side of the last inequality with \(\mu^x y_{d^x-1}y_{d^x} + (2^{\rho}-\mu^x-c_{d^x}) y_{d^x}\) whenever \(2^{\rho}-c_{d^x} > \mu^x\). Handling constraints, despite being possible, is non-trivial for QA; it may introduce exceedingly large coefficients (see discussion in ), and could be contradictory to the purpose of this work. As we presented earlier in [\[binToSpin\]](#binToSpin){reference-type="eqref" reference="binToSpin"} and [\[int2SpinEncoding\]](#int2SpinEncoding){reference-type="eqref" reference="int2SpinEncoding"}, we may directly encode integer variables into spin variables. Moreover, by the fact that in the bounded-coefficient encoding \(\sum_{j=1}^{d^{x}} c_j = \kappa^x\), the encoding of each integer variable into spin variables would be \[\label{intToSpin_BC} x = \frac{1}{2} \l ( \kappa^x + \sum_{j=1}^{d^{x}} c_j^{x} s_j^{x} \r).\] In the next section, where we try to find the upper bound on the coefficients of the encoding, i.e., \(\mu^x\), we use the integer-to-spin transformation.
# Finding the Upper Bound on the Coefficients of the Encoding {#sec:findingUB}
We explained earlier, in Section [1](#sec:intro){reference-type="ref" reference="sec:intro"}, that the QA scales the coefficients of an Ising model to a limited range and has low precision. We also mentioned that resilience to noise could be helpful in overcoming the issue of precision. We propose that restricting the range of the coefficients of the problem could improve resilience. To this end, we wish to find upper bounds on the coefficients of the encoding, i.e., \(\mu\), such that after the encoding of the integer problem [\[UIQP\]](#UIQP){reference-type="eqref" reference="UIQP"} to an Ising model [\[Ising_def\]](#Ising_def){reference-type="eqref" reference="Ising_def"}, the ratio of the smallest local field in magnitude to the largest one exceeds a threshold, i.e., \[\label{localFieldsRatio} \frac{ \min_i |h_i|}{\max_i |h_i|} \ge \epsilon_l.\] Similarly, we wish to bound the ratio of the smallest coupler to the largest one: \[\label{couplerRatio} \frac{ \min_{i,j} |J_{ij} | }{\max_{i,j} |J_{ij} | } \ge \epsilon_c.\] Alternatively, one may desire to have the local fields biases and couplings' strengths well-separated to avoid the effect of noise, or in other words, having \[\min_{i, j} \l | h_i-h_j \r | \ge \epsilon_l,\] and \[\min_{\substack{i,j \\ l,k}} \l | J_{ij}-J_{lk} \r | \ge \epsilon_c,\] could be desirable. In this paper, however, we focus merely on inequalities [\[localFieldsRatio\]](#localFieldsRatio){reference-type="eqref" reference="localFieldsRatio"} and [\[couplerRatio\]](#couplerRatio){reference-type="eqref" reference="couplerRatio"}. Suppose we have a quadratic function \[\label{UIQP_obj} f(x)= x^t Q x + q^t x,\] where \(Q\) is symmetric and the domain of \(f(x)\) is \(x=\l [x_1, x_2, \ldots, x_n \r]^t\) for \(x_i \in \{0, \ 1, \ \ldots, \ \kappa^{x_i}\}\). Note that this is the function which we aim to minimize in [\[UIQP\]](#UIQP){reference-type="eqref" reference="UIQP"}. Also, note that since \(x_i x_j = x_j x_i\) for two integers \(x_i\) and \(x_j\), we may substitute \(Q\) with \(\frac{1}{2} (Q +Q^t)\) when \(Q\) is not symmetric. Of course, everything presented here is under the assumption that the coefficients of the problem (before any encoding) respect inequalities [\[localFieldsRatio\]](#localFieldsRatio){reference-type="eqref" reference="localFieldsRatio"} and [\[couplerRatio\]](#couplerRatio){reference-type="eqref" reference="couplerRatio"}, i.e., unary encoding satisfies them. Let us define the encoding matrix denoted by \(C\) as \[\label{encdMatrix} C= \begin{bmatrix} c^{x_1}_1& c^{x_1}_2 & \dots & c^{x_1}_{d^{x_1}} & 0 & 0 & \dots & 0 & 0 & \dots & \dots & \dots & 0 & 0& 0 & \dots & 0 \\ 0 & 0 & \dots & 0& c^{x_2}_1& c^{x_2}_2 & \dots & c^{x_2}_{d^{x_2}} & 0& \dots & \dots & \dots & 0 & 0& 0 & \dots & 0 \\ \vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots & \ddots & \ddots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & \dots & 0& 0 & 0 & \dots & 0& 0& \dots & \dots & \dots & 0 & c^{x_n}_1& c^{x_n}_2 & \dots & c^{x_n}_{d^{x_n}} \end{bmatrix},\] and the vector of spin variables as \[s = \begin{bmatrix} s^{x_1}_1& s^{x_1}_2 & \dots & s^{x_1}_{d^{x_1}} &s^{x_2}_1& s^{x_2}_2 & \dots & s^{x_2}_{d^{x_2}} & \dots & \dots & \dots & s^{x_n}_1& s^{x_n}_2 & \dots & s^{x_n}_{d^{x_n}} \end{bmatrix}^t.\] Using [\[intToSpin_BC\]](#intToSpin_BC){reference-type="eqref" reference="intToSpin_BC"}, we conclude that the encoding of integer variables, \(x\), into spin variables, \(s\), refers to the following substitution: \[\label{intToSpin_vecForm} x = \frac{1}{2} \l ( \kappa + C s \r),\] where \(\kappa = \l[\kappa^{x_1}, \ \kappa^{x_2}, \ \ldots, \ \kappa^{x_n} \r]^t\) is the vector of all of the upper bounds on the integer variables. Using this substitution in [\[UIQP_obj\]](#UIQP_obj){reference-type="eqref" reference="UIQP_obj"}, we get \[\begin{aligned}
\label{IsingObjFun} f(s) = \frac{1}{4} s^t \l ( C^t Q C-\Delta ( \delta(C^t Q C)) \r) s & + & \frac{1}{2} \l( C^t Q \kappa + C^t q \r)^t s \notag \\ &+& \frac{1}{4} \l( \kappa ^t Q \kappa + e^t \Delta ( \delta(C^t Q C)) e + 2 q^t \kappa \r). \end{aligned}\] We are excluding the diagonal entries of \(C^t QC\) in the quadratic term because they correspond to the square of spin variables, i.e., \(s_i^2\), and \(s_i^2\) is a constant equal to one. The local field bias corresponding to each spin variable and the coupling strength corresponding to each pair of spin variables, derived in [\[IsingObjFun\]](#IsingObjFun){reference-type="eqref" reference="IsingObjFun"}, is summarized below: \[\begin{aligned}
s_j^{x_i} &: & \frac{1}{2} \l [ Q \kappa + q \r ]_i c_j^{x_i} = \frac{1}{2} \l(q_i + \sum_{k=1}^{n} Q_{ik} \kappa_k \r) c_j^{x_i}, \label{hint0} \\ s_k^{x_i}s_l^{x_i} &: & \frac{ \l(Q_{ii} c_k^{x_i} c_l^{x_i}\r) }{2} \quad \text{for } \ k,l \in \{1, \ldots, d^{x_i}\} \ \text{ and } \ k<l, \label{hint1}\\ s_k^{x_i}s_l^{x_j} &: & \frac{ \l(Q_{ij} c_k^{x_i} c_l^{x_j}\r) }{2} \quad \text{for }\ k \in \{1, \ldots, d^{x_i}\} ,\ l \in \{1, \ldots, d^{x_j}\}, \ \text{ and } \ i,j \in \{1, \ldots, n\}: i<j, \label{hint2} \ \ \ \end{aligned}\] where \(\l [ Q \kappa + q \r ]_i\) in the first line denote the \(i\)-th entry of vector \(Q \kappa + q\), as expected, and [\[hint1\]](#hint1){reference-type="eqref" reference="hint1"} and [\[hint2\]](#hint2){reference-type="eqref" reference="hint2"} incorporate the fact that \(s_k^{x_i}s_l^{x_i} = s_l^{x_i}s_k^{x_i}\) and \(s_k^{x_i}s_l^{x_j}= s_l^{x_j}s_k^{x_i}\). Since all of the local fields biases and couplings' strengths share the \(\frac{1}{2}\) factor and after rescaling our concern is their ratio, we will drop the \(\frac{1}{2}\) factor for our calculations in the rest of this section. From the previous section and the derivation of the bounded-coefficient encoding, we conclude that the smallest coefficient used in the encoding is 1; hence, the minimum absolute value of the local fields biases is \[\notag m_l = \min_{i} \left \{ \ \l | \l [ Q \kappa + q \r ]_i \r | \ \right \},\] and the minimum absolute value of couplings' strengths is \[\notag m_c = \min_{i,j} \left \{ \ \l | Q_{ii} \r|, \l| Q_{ij} \r| \ \right \}.\] The maximum absolute value of the local fields biases among all spin variables corresponding to an integer variable \(x_i\) occurs at \(\mu^{x_i}\)---to be found---for each \(i\in \{1,\ldots, n\}\). Therefore, by [\[localFieldsRatio\]](#localFieldsRatio){reference-type="eqref" reference="localFieldsRatio"}, we wish to have: \[\begin{aligned}
\label{LFratio} \frac{m_l} { \l | \ \l [ Q \kappa + q \r ]_i \ \r| \mu^{x_i} } \ge \epsilon_l. \end{aligned}\] Similarly, [\[couplerRatio\]](#couplerRatio){reference-type="eqref" reference="couplerRatio"} enforces the following conditions for the couplers: \[\label{Cratio} \frac{m_c} { |Q_{ii}| (\mu^{x_i})^2 } \ge \epsilon_c, \ \text{and} \ \frac{m_c} { |Q_{ij}| \mu^{x_i}\mu^{x_j} } \ge \epsilon_c.\] Using [\[LFratio\]](#LFratio){reference-type="eqref" reference="LFratio"} and [\[Cratio\]](#Cratio){reference-type="eqref" reference="Cratio"}, we wish to obtain \(\mu^{x_i}\)'s that satisfy the following set of inequalities: \[\begin{aligned}
\label{muIneqList} \mu^{x_i} & \le& \frac{m_l} { \l | \l [ Q \kappa + q \r ]_i \r| \epsilon_l } \label{muIneqList_1} \\ \mu^{x_i} &\le & \sqrt{\frac{m_c}{|Q_{ii} |\epsilon_c} }, \label{muIneqList_2} \\ \mu^{x_i} \mu^{x_j} &\le & \frac{m_c}{|Q_{ij}| \epsilon_c}.\label{muIneqList_3} \end{aligned}\] We may solve a feasibility problem to find a solution to the above set of inequalities, i.e., an optimization or auxiliary objective function along with these inequalities. A well-justified objective function could be \(\max \ \min_i \{ \mu^{x_i} \}\). Solving such problems is normally costly because [\[muIneqList_3\]](#muIneqList_3){reference-type="eqref" reference="muIneqList_3"} is non-convex and \(\mu^{x_i}\)'s are required to be integers. Alternatively, we present an algorithm below that heuristically finds \(\mu^{x_i}\)'s. Note that any set of \(\mu^{x_i}\)'s satisfying the above inequalities will guarantee that [\[localFieldsRatio\]](#localFieldsRatio){reference-type="eqref" reference="localFieldsRatio"} and [\[couplerRatio\]](#couplerRatio){reference-type="eqref" reference="couplerRatio"} hold. Our proposed algorithm for finding \(\mu^{x_i}\)'s first initializes \(\mu^{x_i}\)'s using inequalities [\[muIneqList_1\]](#muIneqList_1){reference-type="eqref" reference="muIneqList_1"} and [\[muIneqList_2\]](#muIneqList_2){reference-type="eqref" reference="muIneqList_2"}. If [\[muIneqList_3\]](#muIneqList_3){reference-type="eqref" reference="muIneqList_3"} is satisfied for all \(i\) and \(j\) at this step, it terminates. Otherwise, it greedily decrease \(\mu^{x_i}\) or \(\mu^{x_j}\) for an \(i\) and \(j\) pair for which the failure happened. In this process, we take \(\frac{\kappa^{x_i} }{\mu^{x_i}}\) as an estimate on the width of the encoding, so when we want to decrease either \(\mu^{x_i}\) or \(\mu^{x_j}\), we choose the one that gives a lower combined width. See Algorithm [\[alg:Find\_\\(mu\\)\_Ising\]](#alg:Find_\(mu\)_Ising){reference-type="ref" reference="alg:Find_\(mu\)_Ising"} for a formal presentation of this algorithm.
[\[table:res_BC\]]{#table:res_BC label="table:res_BC"}
[\[table:res_binary\]]{#table:res_binary label="table:res_binary"}
Note that in, it is suggested to set \(\epsilon = \frac{2}{J_{\max} }\), where \(J_{\max}\) is the largest \(J_{ij}\). Here, our approach slightly differs. We assume that we know the precision of the machine, and we derive the upper bound on the coefficients of our encoding to accommodate this restriction. The precision of the current D-Wave machine is in order of \(10^{-2}\), which motivates our choice of \(\epsilon = 10^{-2}\).
# Conclusion and Discussion {#sec:conclusion}
In this paper, we presented an encoding to represent an unconstrained integer quadratic programming problem as an Ising model. To deal with the low precision of quantum annealers, we suggested bounding the value of the coefficients in the encoding. This restricts the range of the local fields biases and couplings' strengths in the derived Ising model, thereby creating Ising models that are more robust against noise after scaling. Resilience is used as a metric for the robustness of the Ising models. We compared bounded-coefficient encoding with the binary encoding. We infer from our results that bounding the coefficients of the encoding, as in the bounded-coefficient encoding, significantly improves the resilience of the model. The drawbacks are that the size of the derived Ising model is larger and it introduces redundancy. In our proposed technique, we forced the ratios of the minimum absolute value of the local fields biases and the couplings' strengths to their respective maximums exceed a certain threshold tolerance that is correlated or equal to the quantum annealer's precision. However, all of our calculations are embedding-free, i.e., they ignore the Chimera graph structure. In theory, having an embedding with equal chain length for all variables and equal chain connectivity for any present quadratic terms will respect our calculation; however, finding such an embedding is nontrivial. Notice that, after the encoding, all spin variables corresponding to an integer variable are connected; additionally, all spin variables corresponding to \(x_i\) and \(x_j\) are connected if the term \(x_ix_j\) appears in the integer formulation; so, the underlying graph of the Ising model is quite dense. Even for our small instances, the resulting Ising models either exceed the size that can be embedded on the current chip, or the chains' lengths differ in orders of magnitude (e.g., minimum and maximum chain lengths in orders 1 and 10, respectively). Our algorithm can be tested on the future generation of the quantum annealers with improved connectivity. According to our results in the previous section, using \(\epsilon_l = \epsilon_c = 0.01\) to find the upper bound on the coefficients of the encoding, the derived Ising model stays robust against noise, with an average resilience above 0.5, and minimum resilience above \(0.1\) for \(\epsilon= 0.005\), i.e., \(\frac{\epsilon_c}{2}\). This could suggest using \(2 \epsilon_{QA}\) as \(\epsilon_l\) and \(\epsilon_c\) for finding the upper bounds on the coefficients of the encoding, where \(\epsilon_{QA}\) is the precision of the quantum annealer. This, however, requires more experimentation and could be an avenue for future study. | {'timestamp': '2017-06-08T02:01:38', 'yymm': '1706', 'arxiv_id': '1706.01945', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01945'} |
# Introduction {#seq1}
Nowadays, Cloud computing allows data owners to use massive data storage and large computation capabilities at a very low costs. Despite these benefits, such a data outsourcing induces important security challenges. Indeed, data owners lose the control over the pieces of information they outsource. To protect data in terms of confidentiality and privacy from unauthorized users as well as from the cloud, one common solution consists in encrypting data. However, if using encryption achieves data confidentiality, it may limit the possible reuse or processing of outsourced data as well as the sharing of data. In this work, we are interested in the sharing of data between different users who have outsourced their data encrypted with their own public keys, i.e. using some asymmetric cryptosystem. Such a kind of problem is referred as proxy re-encryption (PRE), where Alice (the delegator or the emitter) wants to share with Bob (the delegate or recipient) some data she previously outsourced encrypted into the cloud (the proxy). When working with asymmetric encryption, the objective of PRE is to securely enable the proxy to re-encrypt Alice's cipher-text, encrypted with her public key, into a cipher-text that can be decrypted with Bob's private key. To do so, one simple PRE solution consists in asking Alice to provide her private key to the proxy. However, this strategy imposes the proxy to be completely trusted and does not work in the case the cloud is considered as semi-honest (i.e., it will not disclose the data but will be curious). Blaze *et al*. proposed the first PRE scheme in such a semi-honest framework. This one is based on the ElGamal cryptosystem and on a set of secret pieces of information, referred as secret re-encryption key, Alice has to send to the proxy so as to make possible the change of the public key encryption (i.e., re-encrypt data with Bob's public key). One main issue of this proposal, remarked by Ateniese *et al*., is that Blaze *et al*.'s scheme is inherently bidirectional, that is to say that the re-encryption key which allows transferring cipher-texts from Alice to Bob, enables the proxy to convert all Bob's cipher-texts under Alice's public key. This is not acceptable for Bob. The main reason of this is that the re-encryption key depends on the delegate (Bob) private key. In order to solve this problem and achieve a unidirectional PRE different approaches have been proposed. The first class of methods relies on classical asymmetric encryption cryptosystems. For instance, take advantage of a quorum-based protocol which stands on distributed proxies, each of them possesses a part of the data of Alice but receive a different re-encryption key independent of Bob private key. However, with this approach, the security of Alice private key is safe as long as some proxies are honest. An alternative, proposed in, works with only one proxy where the re-encryption key provided by Alice is split into two parts, one for the proxy and the other for Bob. Unfortunately, with, the data of Alice, she encrypted with her public-key are turned into symmetrically encrypted data and not asymmetrically with the public key of Bob. The second class regroups methods referred as identity-based proxy re-encryption (IBPRE) and was introduced by Green and Ateniese. Such a method mixes PRE with identity-based cryptography (IBC). In IBC, the public encryption key of one user is derived from his identity (e.g., his email address); by combining it with PRE, the emitter and the proxy just need to know the delegates' identities instead of verifying their certificates. Basically, the unidirectional propriety is achieves due to the fact the re-encryption key depends on the identity of the delegate. However, it must be known that IB-PRE suffers of the key-escrow issue (see for more details). Most of these schemes also rely on cryptosystems which are based on bilinear pairing, an application considered as a very expensive in terms of computation complexity compared to modular multiplication or exponentiation. To overcome this issue, Deng *et al*. proposed an asymmetric cross-cryptosystem re-encryption scheme instead of pairing. Beyond, if the above approaches allow one user to share data with another one, they do not make possible the processing of encrypted data by the cloud or proxy. This capacity is usually achieved with the help of homomorphic cryptosystems. With these ones, one can perform operations onto encrypted data with the guarantee that the decrypted result equals the one carried out onto un-encrypted data. The first homomorphic based PRE attempt has been proposed by Bresson *et al*. in, using the Paillier cryptosystem. However, even though their solution makes possible data sharing, it cannot be seen as a pure proxy re-encryption scheme. Indeed, data are not re-encrypted with the public key of the delegate. If this one wants to ask the cloud to process the data he receives from Alice, he has: i) first to download Alice data, ii) decrypt them based on some secret pieces of information provided by Alice; iii) re-encrypt them with his public key and send them back to the cloud. There is thus still a need for a homomorphic based PRE. In this work, we propose the first homomorphic proxy re-encryption scheme which does not require the delegate to re-upload the data another user has shared with him. It is based on the Paillier cryptosystem. It can be roughly summarized as follows. Bob and Alice agree on a secret key; key Alice sends Paillier encrypted to the cloud. The cloud uses this key so as to generate a Paillier encrypted random sequence with the help of a secure linear congruential generator (SLCG) we propose and which works in the Paillier encrypted domain. All computations are conducted by the cloud server. This SLCG provides a sequence of Paillier encrypted random numbers. Based on a fast and new solution we propose so as to compute the difference in-between Paillier encrypted data, the cloud: i) computes in clear the difference between this encrypted random sequence and the encrypted data of Alice and, ii) encrypts this sequence of differences with the public key of Bob. Then, Bob just has to ask the cloud to remove the noise from the encrypted data in order to get access to the data Alice wants to share with him and process them in an outsourced manner if he wants. The rest of the paper is organized as follow. In Section [2](#seq2){reference-type="ref" reference="seq2"}, we come back on the definition of Paillier cryptosystem and show how to use it in order to: i) quickly compute the difference between Paillier encrypted data; and ii) implement a secure linear congruential generator so as to generate an encrypted random sequence. Section [3](#seq3){reference-type="ref" reference="seq3"} describes the overall architecture of our Homomorphic PRE solution (HPRE) in the case of the sharing of images. Performance of the proposed solution is given in Section [4](#seq4){reference-type="ref" reference="seq4"}. Conclusions are given in Section [5](#seq6){reference-type="ref" reference="seq6"}.
# Processing Paillier Encrypted Data {#seq2}
In this section, we first introduce the Paillier cryptosystem as well as a new way to compute the difference between Paillier encrypted data before presenting a secure linear congruential generator (LCG) implemented in the Paillier encrypted domain so as to generate an encrypted pseudo random sequence of integers.
## Paillier cryptosystem {#sseq2:1}
We opted for the asymmetric Paillier cryptosystem because of its additive homomorphic property. In this work, we use a fast version of it defined as follows. Let \(((g,K_p), K_s)\) be the public/private key pair, such as: \[K_p = pq \quad and \quad K_s= (p-1)(q-1) \label{eq1}\] where \(p\) and \(q\) are two large prime integers. \(\mathbb{Z}_{K_p}= \{0, 1..., K_p-1\}\) and \(\mathbb{Z}_{K_p}^*\) denotes the integers that have multiplicative inverses modulo \(K_p\). We select \(g\in\mathbb{Z}_{K_p^2}^*\) such as: \[\frac{g^{K_s}-1 \, mod\, K_p^2}{K_p} \in \mathbb{Z}_{K_p}^* \label{eq2}\] The Paillier encryption of a plain-text \(m\in\mathbb{Z}_{K_p}\) into the cipher-text \(c\in\mathbb{Z}_{K_p^2}^*\) using the public key \(K_p\) is given by \[c= E[m,r]= g^m r^{K_p} \mod K_p^2 \label{eq3}\] where \(r \in \mathbb{Z}_{K_p}^*\) is a random integer associated to \(m\) making the Paillier cryptosystem probabilistic or semantically secure. More clearly, depending on the value of \(r\), the encryption of the same plain-text message will yield to different cipher-texts even though the public encryption key is the same. Notice that it is possible to get a fast version of [\[eq3\]](#eq3){reference-type="eqref" reference="eq3"} by fixing \(g=1+K_p\) without reducing the algorithm security. By doing so, the encryption of \(m\) into \(c\) requires only one modular exponentiation and two modular multiplications \[c=E[m,r]=(1+mK_p)r^{K_p} \mod K_p^2 \label{eq4}\] As we will see in section [2.2](#sseq2:2){reference-type="ref" reference="sseq2:2"}, this property will be of importance for the computation of the difference between Paillier encrypted data.\
Based on the assumption \(g=1+K_p\), the decryption of \(c\) using the private Key \(K_s\) is such as \[m=\frac{(c^{K_s}-1)K_s^{-1} \mod K_p^2}{K_p} \mod K_p \label{eq5}\] If we consider two plain-texts \(m_1\) and \(m_2\), the additive homomorphic property of the Paillier cryptosystem allows linear operations on encrypted data like addition and multiplication, ensuring that \[E[m_1,r_1]E[m_2,r_2] = E[m_1+m_2, r_1r_2] \label{eq6}\] \[E[m_1,r_1]^{m_2} = E[m_1 m_2, r_1^{m_2}] \label{eq7}\]
## Computing the difference in-between encrypted data {#sseq2:2}
In this work, we propose a solution that allows the calculation by one server of the difference between two Paillier encrypted data. More clearly if \(a\) and \(b\) are two integers, we want to compute their difference \(a-b\) from their encrypted versions. Let us consider a user-server relationship where the server has two cipher-texts \(E_{K_p}[a,r]\) and \(E_{K_p}[b,r]\) encrypted by the user. It is important to notice that to make such computation possible; the two cipher-texts have to be encrypted with the same random value \(r\). Under this constraint, one can directly derive the difference \(d\) between \(a\) and \(b\) from \(E_{K_p}[a,r]\) and \(E_{K_p}[b,r]\) by taking advantage of the fast Paillier cryptosystem assumption, i.e. \(g=1+K_p\), as follows \[\begin{array}{ccc} d & = & D(a,b) =D^e(E_{K_p}[a, r], E_{K_p}[b, r]) \\[.3cm] & = & \frac{E_{K_p}[a,r]E_{K_p}[b,r]^{-1}-1 \mod K_p^2}{K_p} \mod K_p \\[.3cm] & = & \frac{g^a r g^{-b}r^{-1}-1 \mod K_p^2}{K_p} \mod K_p \\[.3cm] & = & \frac{g^{a-b}-1 \mod K_p^2}{K_p} \mod K_p \\[.3cm] d & = & a-b \mod K_p \end{array}\] \[\label{eq8}\] where \(D\) and \(D^e\) denote the two functions that allows computing the difference \(d\) in the clear and Paillier encrypted domain, respectively. Notice that knowing the difference \(d\) between \(a\) and \(b\) gives no clues about the values of \(a\) and \(b\), respectively.
## Secure Linear Congruential Generator {#sseq2:3}
As stated in the introduction, our HPRE scheme will require the cloud to securely generate a pseudo random sequence that is to say a Paillier encrypted random sequence of integers. The generator we propose to secure is LCG (Linear Congruential Generator). This one is based on congruence and a linear functions; functions that can be easily implemented in the Paillier encrypted domain. In the clear domain, LCG works as follows \[X_{n+1} = a X_n + c \mod m \label{eq9}\] where: \(X_n\) is the \(n^{th}\) random integer value of the LCG sequence; \(a\) is a multiplier; \(c\) is an increment; \(m\) is the modulo; and, \(X_0\) the initial term, also called the seed or the secret LCG key, one needs to know so as to re-generate a random sequence. The security of the LCG is based on the seed \(X_0\). The knowledge of the parameters \(a\), \(c\) and \(m\) does not endanger its security. This random generator can be implemented into the Paillier encrypted domain, i.e. turned into a Secure LCG (SLCG), so as to generate an encrypted random sequence of integers (i.e. \(\{E[X_n,r_n ]\}_{n=0...N-1}\)) in the following way: \[E[X_{n+1},r_{n+1}]=E[X_n,r_n ]^a E[c,r_c] =E[a X_n+c,r_n^ar_c] \label{eq10}\] under the constraint however that \(m\) equals the user Paillier public key \(K_p\), (i.e., \(m = K_p\), see [\[eq1\]](#eq1){reference-type="eqref" reference="eq1"}). If the increment as well as all terms of the sequence are encrypted (including the LCG seed) that is not the case of the multiplier \(a\). However, this does not reduce the security of our system as the parameter \(a\) is not supposed to be secret. It is important to notice that, in our SLCG, a recursive relation exists between the random integers \(r_n\) which ensure the semantic security of the Paillier cryptosystem. Derived from [\[eq10\]](#eq10){reference-type="eqref" reference="eq10"}, this one is such as: \[r_{n+1}= r_n^a r_c \label{eq11}\] where \(r_c\) is the random variable used to encrypt the increment. \(r_0\) is the random value associated to the seed \(X_0\). This recursive relationship will be considered in Section [3.2](#sseq3:2){reference-type="ref" reference="sseq3:2"} so as to allow data exchange between two different users.
# Sharing outsourced encrypted data {#seq3}
In this Section, we first refine the data exchange framework we consider and its basic security assumptions. We then present our Homomorphic based Proxy Re-Encryption scheme (HPRE).
## Data exchange scenario in outsourced environment {#sseq3:1}
Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} illustrates the general data exchange framework we consider where a data owner (the emitter or the delegator) has *a priori* stored his data into a public cloud in an asymmetrically encrypted form; data he wants to share with another user (the recipient or the delegate). We further assume a semi-honest cloud server. This one honestly stores encrypted data uploaded by the users and responds to their requests. If the server does not disclose data to any parties who fail to prove ownership or access rights, it is however curious and may try to infer information about the content of users' data or about their private keys. One last assumption is that all communications between the server and the users are protected with the help of the Paillier cryptosystem. Eavesdroppers cannot infer messages being transmitted. As stated previously, our objective is to allow them to share some data under the constraint the delegator does not have to download his data, re-encrypt them with the public key of the delegate and upload them into the cloud. We also want this process conducted by the cloud (proxy) without giving the delegator private key as well as with very few communications in-between the delegator, the proxy and the delegate. In our idea, if one user wants to share data with several users at once, all of them will have to agree on a single secret with the delegator.
## Secure data exchange between users {#sseq3:2}
Let us thus consider that Alice (the delegator) wants to share with Bob (the delegate) a set of data she is the owner of. These data could be a set of integer values like for instance a gray-scale image \(I\), the \(N\) pixels of which \(I=\{I_i\}_{i=0..N-1}\) are encoded on \(b\) bits. As stated previously, it is assumed that Alice has already outsourced an image into the cloud by Paillier encrypting its pixels independently with her public key \(K_{p1}\), such as (see in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} -- Data outsourcing step) \[I_i^e=E_{K_{p1}}[I_i,r_i] \label{eq12}\] where \(r_i\) is the random value associated to the \(i^{th}\) pixel \(I_i\) of \(I\), \(I_i^e\) is the encrypted version of \(I_i\). As we will see in the sequel, our HPRE procedure imposes a constraint on the way Alice generates the random values \(\{r_i\}_{i=0..N-1}\). These ones should satisfy [\[eq11\]](#eq11){reference-type="eqref" reference="eq11"}, that is to say that for one file Alice stores into the cloud, she has to memorize the first random value \(r_c\) and \(r_0\) she used when she encrypted the first pixel of her image (or of any files she stored), \(I_0^e=E_{K_{p1}}[I_0,r_0]\). In order to share this encrypted image with Bob, the public Paillier encryption key of whom is \(K_{p2}\), we propose the following HPRE procedure also depicted in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}
1. **User agreement for data exchange**-In this first tsep, Bob and Alice have to agree on the exchange by defining the LCG parameters, in other words: the secret key \(X_0\), the multiplier \(a\) and the increment \(c\). Let us recall that knowing \(c\) and \(a\) is not critical from a security point of view (see Section [2.3](#sseq2:3){reference-type="ref" reference="sseq2:3"}).
2. **Secret random sequence generation**-Alice encrypts \(X_0\) and \(c\) under her public key \(K_{p1}\): \(E_{K_{p1}}[X_0,r_0]\) \(E_{K_{p1}}[c,r_c]\), and sends them to the cloud. Notice that \(X_0\) is encrypted with the same random integer \(r_0\) Alice used to encrypt the first pixel of her image (see above). She also sends the multiplier \(a\). Based on these pieces of information, the cloud generates the secret random sequence \(X^e=\{X_i^e=E_{K_{p1}}[X_i,r_i]\}_{i=0..N-1}\) using [\[eq10\]](#eq10){reference-type="eqref" reference="eq10"}.
3. **Data encryption for the delegator**-This procedure relies on different stages: i) the computation of differences between the encrypted data of Alice \((I^e)\) and the secret random sequence \((X^e)\); ii) the encryption of this differences with the public key of Bob \(K_{p2}\).
1. *Difference computation*-since \(X_i^e\) and \(I_i^e\) have been encrypted with the same public key \(K_{p1}\) and the same random values \(r_i\) (see above), the cloud computes their difference \(D_i\) as exposed in see Section [2.2](#sseq2:2){reference-type="ref" reference="sseq2:2"}, that is to say \[\begin{small} \begin{array}{ccc} D_i & = &D(X_i,I_i )=D^e(E_{K_{p1}}[X_i,r_i],E_{K_{p1}}[I_i,r_i ]) \\ • & = & X_i-I_i \mod K_{p1} \end{array} \end{small} \label{eq13}\] Even though the cloud knows \(D=\{D_i\}_{i=0...N-1}\), it cannot deduce the value of \(I_i\) and \(X_i\).
2. *Data encryption for the delegator*-From this stand point, one may think the cloud just has to encrypt \(D\) with the public key of Bob, \(K_{p2}\), and then remove the noise so as so to give him access to the data. This is possible under the constraint \(D_i \mod K_{p1} = D_i \mod K_{p2}\) which is achieved when \(0<D_i<min(K_{p1}, K_{p2})\). Unfortunately, this constraint is hard to satisfy because of the SLCG the output amplitude of which can not be controlled simply. To overcome this issue, our HPRE includes a \"noise refreshment procedure\" (see Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}) before encrypting the data with the public key of Bob.
- Noise refreshment
To refresh the noise, Bob first generates on his side the sequence \(\{X_i\}_{i=0..N-1}\), using an LCG parameterized as the SLGC of the cloud. He also produces a second noise \(\{\beta_i\}_{i=0..N-1}\) such as: \[2^b-1<\beta_i<\min(K_{p1},K_{p2}) \label{eq14}\] where \(b\) is the number of bits on which is encoded the pixel values of the image of Alice. Under such a constraint: we ensure: \(\beta_i \mod K_{p1}= \beta_i \mod K_{p2}\) and \(\beta_i-I_i \mod K_{p1}= \beta_i-I_i \mod K_{p2}\). Then Bob sends to the cloud \(\{E_{K_{p2}}[\beta_i,r_i']\}_{i=0...N-1}\) and \(\{\alpha_i=\beta_i-X_i \mod K_{p1}\}_{i=0..N-1}\). Where \(r_i'\) is a random value defined by Bob. On its side, in order to remove the noise \(\{X_i\}_{i=0...N-1 }\), the cloud computes \[G_i=\alpha_i+D_i \mod K_{p1}=\beta_i-I_i \mod K_{p1} \label{eq15}\] Then it encrypts \(\{G_i\}_{i=0..N-1}\) with the public key of Bob \[\{E_{K_{p2}}[G_i,r_i'']=E_{K_{p2}}[\beta_i-I_i,r_i'']\}_{i=0..N-1} \label{eq16}\] Finally, in order to remove the noise \(\beta_i\) from of the data of Bob, the server computes \[E_{K_{p2}}[I_i,r_i' r_{i}^{''-1}]=E_{K_{p2}}[\beta_i,r_i']E_{K_{p2}}[\beta_i-I_i,r_i'']^{-1} \label{eq17}\]
At the end of this procedure, Bob has on the cloud the image of Alice encrypted with his own public key.
As depicted, this system allows the data exchange between Alice and Bob, without extra-communication between the cloud and Alice, and the downloading of data by Bob. It is also possible to notice that the access to the shared data is based on the knowledge of the secret SLCG key \(X_0\) generated by Alice in agreement with Bob. Because our scheme is based on homormophic encryption, data can be by next processed by the cloud without endangering data confidentiality.
:::
# Experimental results {#seq4}
The previous solution was experimented in the case of the sharing of uncompressed images between two users. These images are issued from the Olivetti Research Laboratory of Cambridge, UK. It contains \(400\) images of \(8\) bit encoded of \(92\times 122\) pixels. Some samples of our image test set are given Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}. These images were encrypted with Paillier public keys of more than \(1024\) bits in order to provide a high level of security. Performance of our scheme are evaluated in terms of storage and computation complexity. Our HPRE was implemented in C/C++ with GMP library and all experiments were conducted using a machine equipped with \(23\)GB RAM running on Ubuntu \(14.04\) LTS.
- Storage complexity:
Assuming that images are Paillier encrypted with a key of \(1024\) bits, one encrypted image needs \(2,7Mo\) so as to be stored into the cloud. For one image the delegator (Alice) outsources, she only has to store on her side the random values \(r_0\) and \(r_c\). However this is not obligation. Indeed, based on the fact she knows both her public and private keys, she just has to download the encrypted seed \(E_{K_{p1}}[X_0,r_0]\) and the encrypted increment \(E_{K_{p1}}[c,r_c]\), to get access to these random values (i.e. \(r_c\) and \(r_0\)). During an image exchange, the delegator sends the encrypted seed \(E_{K_{p1}} [X_0,r_0]\), the encrypted increment \(E_{K_{p1}}[c,r_c]\) and the multiplier \(a\). This amount of data is bounded by \(O(log_2(K_{p1}^2))\). For a key of \(1024\) bits, it is closed to \(2048\) bits. On its side, the delegate (Bob) has to store \(X_0\), the secret key of the LCG, but only for one session of data exchange.
- Computation complexity:
On the delegator side, the computation complexity is limited to the encryption of the SLCG parameters (i.e. \(X_0\), \(c\)). Such a complexity is independent of the image's size. Regarding the cloud, this one has to compute: the secret random sequence, compute the difference between the encrypted date of Alice with this random sequence, refresh the noise based on the inputs of Bob, encrypt the result with the public key of Bob and finally remove the noise. For an image of \(N\) pixels, the secret random sequence generation is equivalent to \(N\) encryptions. It is the same for the computation of the differences \(\{D_i\}_{i=0...N-1}\). As described above, the noise refreshment procedure consists in modular additions. We consider its complexity negligible compared to encryption operations. The last step, the encryption of the differences \(\{G_i\}_{i=0...N-1}\) is made of \(N\) encryptions. As a consequence, the computation complexity for the cloud is bounded by \(O(3\times N)\) encryptions. The delegate computation complexity is attached to the noise refreshment procedure. He has to generate a LCG noise (i.e. \(\{X_i\}_{i=0...N-1}\)), a task the complexity of which is negligible compared to the \(N\) encryptions of the second noise (i.e. \(\{\beta_i\}_{i=0...N-1}\)) he also produces and that he next sends to the cloud. The computation complexity of the delegate is thus of \(N\) encryptions. We provide in Table [\[tab1\]](#tab1){reference-type="ref" reference="tab1"} the amount of data that each entities has to store as well as the computation time required in the case of sharing images of our data set. Our HPRE scheme takes about \(1'30\) minutes so as to share an image with a standard computer.
# Conclusion {#seq6}
In this paper, we proposed the first homomorphic proxy re-encryption scheme. Its originality stands on a solution we propose so as to compute the difference of data encrypted with the fast version of the Paillier cryptosystem. It takes also advantage of a secure linear congruential generator we implemented in the Paillier encrypted domain. This one drastically reduces the computation complexity of the cloud and delegator. Furthermore, this solution doesn't need extra communication between the cloud and the delegator, i.e. the data owner. Moreover, since the data are homomorphically encrypted, it is possible to process outsourced data while ensuring their confidentiality. Our HPRE was implemented in the case of the sharing of uncompressed images stored in the cloud showing good time computation performance. Our scheme is not limited to images and can be used with any kinds of data. | {'timestamp': '2017-06-07T02:06:39', 'yymm': '1706', 'arxiv_id': '1706.01756', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01756'} |
null | null |
# Introduction {#s:intro}
Let \(G\) be a reductive algebraic group defined over an algebraically closed field \(K\) of characteristic \(p\) (possibly equal to 0), \(C\) be a conjugacy class of \(G\) and \(u\) be a positive integer. In 2007 Guralnick proved the following result:
In this paper, we concentrate our attention to the case where \(G\) is connected and \(C=\{1\}\) is the trivial conjugacy class of \(G\). For a positive integer \(u\), we let \[G_{[u]}=\{g \in G: g^u=1\}\] be the subvariety of \(G\) consisting of elements of \(G\) of order dividing \(u\) and set \(j_u(G)= \dim G_{[u]}\). We also let \(d_u(G)\) be the minimal dimension of a centralizer in \(G\) of an element of \(G\) of order dividing \(u\). We are merely interested in determining \(j_u(G)\) for every positive positive integer \(u\) (when \(G\) is a simple algebraic group).\
For completeness, we begin by proving Guralnick's result in the case where \(G\) is connected and \(C=\{1\}\). In the statement below, given \(g \in G\), we let \(g^G\) denote the conjugacy class of \(g\) in \(G\).
Given a simple algebraic group \(G\) defined over an algebraically closed field \(K\) of characteristic \(p\), we denote by \(G_{s.c.}\) (respectively, \(G_{a.}\)) the simple algebraic group over \(K\) of simply connected (respectively, adjoint) type having the same Lie type and Lie rank as \(G\). We prove the following result partially proved in:
In, given a positive integer \(u\) and a simple algebraic group \(G\) defined over an algebraically closed field \(K\) of characteristic \(p\), Lawther gives a lower bound for \(d_u(G)\). Moreover, he proves that this lower bound is equal to \(d_u(G)\) in the case where \(G\) is of adjoint type. We establish the following result completing Lawther's result. Recall that the Coxeter number of a simple algebraic group is defined to be the quotient of the number of roots in the root system associated to \(G\) by the number of such roots which are simple. Concretely, \(h=\ell+1\), \(2\ell\), \(2\ell\), \(2\ell-2\), \(6\), \(12\), \(12\), \(18\) or \(30\) according respectively as \(G\) is of type \(A_\ell\), \(B_\ell\), \(C_\ell\), \(D_\ell\), \(G_2\), \(F_4\), \(E_6\). \(E_7\) or \(E_8\).
We note that for \(G\) of exceptional type the only cases where the simply connected and the adjoint groups are not abstractly isomorphic is when \(G\) is of type \(E_6\) or \(E_7\) and \(p\neq 3\) or \(p\neq 2\) respectively.\
Recall that if a simple algebraic group \(G\) over an algebraically closed field \(K\) of characteristic \(p\) is not of simply connected type nor of adjoint type then either \(G\) is abstractly isomorphic to \({\rm SL}_n(K)/C\) where \(C\leqslant Z({\rm SL}_n(K))\), or \(G\) is of type \(D_{\ell}\), \(p \neq 2\) and \(G\) is abstractly isomorphic to \({\rm SO}_{2\ell}(K)\) or a half-spin group \({\rm HSpin}_{2\ell}(K)\) where \(\ell\) is even in the latter case. For simple algebraic groups that are neither of simply connected nor adjoint type we obtain the following result.
In Lawther showed that if \(G\) is a simple algebraic group of rank \(\ell\) with Coxeter number \(h\) over an algebraically closed field and \(u\) is a positive integer with \(u\geq h\) then \(d_u(G)=\ell\). As a corollary of Theorems [\[t:blawther\]](#t:blawther){reference-type="ref" reference="t:blawther"} and [\[t:duotsca\]](#t:duotsca){reference-type="ref" reference="t:duotsca"} we obtain the following result:
Given a simple algebraic group \(G\) over an algebraically closed field, we also derive a further property of \[d_u(G): \begin{array}{l} \mathbb{N}\rightarrow \mathbb{N}\\ u \mapsto d_u(G). \end{array}\]
We now apply our results to the study of finite quasisimple images of triangle groups. Recall that a finite group is quasisimple if it is perfect and the quotient by its centre is simple.\
Every finite quasisimple group being 2-generated, given a finite quasisimple group \(G_0\) and a triple \((a,b,c)\) of positive integers, a natural question to consider is whether \(G_0\) can be generated by two elements of orders dividing respectively \(a\) and \(b\) and having product of order dividing \(c\). A finite group generated by two such elements is called an \((a,b,c)\)-group. Equivalently, an \((a,b,c)\)-group is a finite quotient of the triangle group \(T=T_{a,b,c}\) with presentation \[T=T_{a,b,c}=\langle x,y,z:x^a=y^b=z^c=xyz=1\rangle.\] When investigating the finite (nonabelian) quasisimple quotients of \(T\), we can assume that \(1/a+1/b+1/c<1\) as otherwise \(T\) is either soluble or \(T\cong T_{2,3,5}\cong{\rm Alt}_5\) (see ). The group \(T\) is then a hyperbolic triangle group. Without loss of generality, we will further assume that \(a \leq b \leq c\) and call \((a,b,c)\) a hyperbolic triple of integers. (Indeed, \(T_{a,b,c}\cong T_{a',b',c'}\) for any permutation \((a',b',c')\) of \((a,b,c)\).)\
Recall that a finite quasisimple group \(G_0\) of Lie type occurs as the derived subgroup of the fixed point group of a simple algebraic group \(G\), when \({\rm char}(K)=p\) is prime, under a Steinberg endomorphism \(F\), i.e. \(G_0=(G^F)'\). We use the standard notation \(G_0=(G^F)'= G(q)\) where \(q=p^r\) for some positive integer \(r\). (We include the possibility that \(G(q)\) is of twisted type.)\
Given a simple algebraic group \(G\) and a triple \((a,b,c)\) of positive integers, we say that \((a,b,c)\) is rigid for \(G\) if the sum \(j_a(G)+j_b(G)+j_c(G)\) is equal to \(2\dim G\). When the latter sum is less (respectively, greater) than \(2\dim G\), we say that \((a,b,c)\) is reducible (respectively, nonrigid) for \(G\).\
In we conjectured the following finiteness result in the rigid case (first formulated for hyperbolic triples \((a,b,c)\) of primes):
In, Larsen, Lubotzky and the author proved that Conjecture [\[c:marionconj\]](#c:marionconj){reference-type="ref" reference="c:marionconj"} holds except possibly if \(p\) divides \(abcd\) where \(d\) is the determinant of the Cartan matrix of \(G\). In a future paper we will make further progress on Conjecture [\[c:marionconj\]](#c:marionconj){reference-type="ref" reference="c:marionconj"} by, in particular, completely settling it in the case where the finite group \((G^F)'\) is simple. A crucial ingredient therein is the classification of the hyperbolic triples \((a,b,c)\) of integers as reducible, rigid or nonrigid for a given simple algebraic group \(G\). We determine this classification below.
As Theorem [\[t:classification\]](#t:classification){reference-type="ref" reference="t:classification"} gives the list of the rigid triples of integers for a given simple algebraic group, it puts Conjecture [\[c:marionconj\]](#c:marionconj){reference-type="ref" reference="c:marionconj"} into a very concrete context. Concerning the reducible case, we prove the following nonexistence result.
Theorem [\[t:classification\]](#t:classification){reference-type="ref" reference="t:classification"} which gives inclusively the list of the reducible triple of integers for a given simple algebraic group is one of the ingredients of the proof of Proposition [\[p:marionred\]](#p:marionred){reference-type="ref" reference="p:marionred"}. As an immediate corollary of Theorem [\[t:classification\]](#t:classification){reference-type="ref" reference="t:classification"} and Proposition [\[p:marionred\]](#p:marionred){reference-type="ref" reference="p:marionred"} we obtain the following result.
Unless otherwise stated, we let \(G\) be a connected reductive algebraic group defined over an algebraically closed field \(K\) of characteristic \(p\) with maximal torus \(T\). If \(T'\) is a torus of \(G\) of dimension \(r\) we sometimes write \(T'=T_r\).\
Given a positive integer \(u\), we let \(G_{[u]}\) be the subvariety of \(G\) consisting of elements of order dividing \(u\), we set \[j_u(G)= \dim G_{[u]}\] and we let \(d_u(G)\) be the minimal dimension of a centralizer in \(G\) of an element of \(G\) of order dividing \(u\).\
For \(G\) simple, we fix some more notation as follows. We let \(\Phi\) be the root system of \(G\) with respect to \(T\) and set \(\Pi=\{\alpha_1,\dots, \alpha_\ell\} \subset \Phi\) to be a set of simple roots of \(G\), where \(\ell\) is the rank of \(G\). We let \(h=|\Phi|/{\rm rank}(G)=|\Phi|/\ell\) be the Coxeter number of \(G\). Recall that \(h=\ell+1\), \(2\ell\), \(2\ell\), \(2\ell-2\), \(6\), \(12\), \(12\), \(18\) or \(30\) according respectively as \(G\) is of type \(A_\ell\), \(B_\ell\), \(C_\ell\), \(D_\ell\), \(G_2\), \(F_4\), \(E_6\). \(E_7\) or \(E_8\). Also we denote by \(G_{s.c.}\) (respectively, \(G_{a.}\)) the simple algebraic group over \(K\) of simply connected (respectively, adjoint) type having the same Lie type and Lie rank as \(G\).\
Unless otherwise stated given a positive integer \(u\), we also write:\
\(u=qv\) where \(q\) and \(v\) are coprime positive integers such that \(q\) is a power of \(p\) and \(p\) does not divide \(v\),\
\(h=zu+e\) where \(z,e\) are nonnegative integers such that \(0\leq e \leq q-1\),\
\(h=\alpha v+\beta\) where \(\alpha,\beta\) are nonnegative integers such that \(0\leq \alpha \leq v-1\),\
\(\alpha=\gamma q+\delta\) where \(\gamma,\delta\) are nonnegative integers such that \(0\leq \delta \leq q-1\).\
An easy check yields \(\gamma=z\). Note that if \(e=0\) then \(\beta=\delta=0\). Furthermore, if \(\beta =0\) then \(\delta=0\) if and only if \(e=0\).\
Finally for an nonnegative integer \(r\), we let \(\epsilon_r \in \{0,1\}\) be \(0\) if \(r\) is even, otherwise \(\epsilon_r=1\), and set \(\sigma_r\in\{0,1\}\) to be 1 if \(r=0\), otherwise \(\sigma_r=0\).\
The outline of the paper is as follows. In §[\[s:precent\]](#s:precent){reference-type="ref" reference="s:precent"} we give some preliminary results on centralizers. In §[\[s:tfr\]](#s:tfr){reference-type="ref" reference="s:tfr"} we give detailed proofs for Propositions [\[p:fcu\]](#p:fcu){reference-type="ref" reference="p:fcu"} and [\[p:ingsimple\]](#p:ingsimple){reference-type="ref" reference="p:ingsimple"}. In §[\[s:ex\]](#s:ex){reference-type="ref" reference="s:ex"}, given a positive integer \(u\), we determine \(d_u(G)\) for \(G\) a simple simply connected algebraic group of exceptional type over an algebraically closed field \(K\). This establishes Theorem [\[t:blawther\]](#t:blawther){reference-type="ref" reference="t:blawther"} for \(G\) of exceptional type. In §[\[s:spin\]](#s:spin){reference-type="ref" reference="s:spin"} we recall some properties of the spin groups and determine when a preimage under the canonical map \({\rm Spin}_n(K) \rightarrow {\rm SO}_n(K)\) of a semisimple element of \({\rm SO}_n(K)\) of a given order is also of that order. In §[\[s:upc\]](#s:upc){reference-type="ref" reference="s:upc"} given a positive integer \(u\), we determine precise upper bounds for \(d_u(G)\) where \(G\) is a simple algebraic group of classical type over an algebraically closed field. In §[\[s:pc\]](#s:pc){reference-type="ref" reference="s:pc"} given a positive integer \(u\), we determine \(d_u(G)\) for \(G\) of classical type, completing the proofs of Theorems [\[t:blawther\]](#t:blawther){reference-type="ref" reference="t:blawther"} and [\[t:duotsca\]](#t:duotsca){reference-type="ref" reference="t:duotsca"}. In §[\[s:decreasing\]](#s:decreasing){reference-type="ref" reference="s:decreasing"} we prove that for a given simple algebraic group \(G\) over an algebraically closed field, \(d_u(G)\), seen as a function of \(u\), is decreasing. This establishes Proposition [\[p:dudecreasing\]](#p:dudecreasing){reference-type="ref" reference="p:dudecreasing"}. In §[\[s:classification\]](#s:classification){reference-type="ref" reference="s:classification"} we classify the reducible and the rigid hyperbolic triples of integers for simple algebraic groups, establishing Theorem [\[t:classification\]](#t:classification){reference-type="ref" reference="t:classification"}. In §[\[s:reducibility\]](#s:reducibility){reference-type="ref" reference="s:reducibility"} we prove Proposition [\[p:marionred\]](#p:marionred){reference-type="ref" reference="p:marionred"}. Finally in §[\[s:tables\]](#s:tables){reference-type="ref" reference="s:tables"} we collect Tables [\[t:asc\]](#t:asc){reference-type="ref" reference="t:asc"}-[\[ta:casestoconsider\]](#ta:casestoconsider){reference-type="ref" reference="ta:casestoconsider"} which appear later in the paper.\ | {'timestamp': '2017-06-07T02:04:32', 'yymm': '1706', 'arxiv_id': '1706.01653', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01653'} |
null | null |
# Introduction {#sec:intro}
The nature of Type Ia supernova (SN Ia) progenitors remains one of the enduring mysteries of astrophysics (for recent reviews, see @hill13a and @maoz14a). For decades, many researchers favored a scenario involving a C/O white dwarf (WD) whose mass approaches the Chandrasekhar limit via stable hydrogen-rich accretion from a non-degenerate companion or in an unstable merger with another C/O WD. Carbon fusion at the center of the WD would then lead to a phase of convective simmering, followed by the birth of a deflagration, a transition to a detonation, and subsequently, a SN Ia explosion (e.g., @kow97 [@plew04a; @seit13b]). However, growing constraints from recent theoretical and observational work have increased persisting doubts that the Chandrasekhar-mass (\(M_{\rm Ch}\)) scenario is responsible for the bulk of SNe Ia (e.g., @leon07 [@sb07; @kerz09a; @rbf09; @kase10; @bloo12; @sp12; @wood13a; @scal14b; @joha16a; @dhaw17a]). Increased attention is being paid to alternative solutions, chief among them the double detonation scenario. In its earliest incarnations, this scenario invoked accretion from a non-degenerate helium-burning star onto a C/O WD, which leads to a \(\sim 0.1 \, M_\sun\) helium shell that ignites, begins to convect, and then detonates. The helium shell detonation then triggers a detonation in the sub-\(M_{\rm Ch}\) C/O core via a direct edge-lit detonation or via shock convergence near the center. However, the helium detonation in the massive shells of these early models produced \(^{56}\)Ni and other iron-group elements in the outer regions of the SN ejecta, which presented problems when compared to observations. In recent years, the realization that stable accretion from helium WD donors yields much smaller helium shells at ignition due to the higher accretion rates, coupled with the problems besetting \(M_{\rm Ch}\) scenarios, motivated a resurgence of double detonation studies focused on the explosion of sub-\(M_{\rm Ch}\) WDs. In parallel work, studies of unstable double WD mergers uncovered the possibility that helium could detonate as it was transferred during the dynamical phase of the merger. This scenario was made even more attractive due to work that showed that including a large nuclear reaction network and realistic C/O pollution in the helium shell drastically reduces the minimum hotspot size and shell mass for helium detonation initiation and propagation. Observational studies have begun to narrow the highly uncertain double WD interaction rate, finding rough agreement with binary population synthesis calculations (e.g., @ruit11 [@toon17a]). A recent observational estimate finds that the rate of double WDs coming into mass transfer contact is \(\sim 10\) times the SN Ia rate. Not all of these binaries necessarily lead to double WD mergers, but introduced the possibility that all double WD systems do indeed merge unstably due to dynamical friction during the initial phases of stable hydrogen-and helium-rich mass transfer. Thus, double WD binaries have the potential to explain all SNe Ia if just \(\sim 10\%\) of double WD mergers lead to SNe Ia via double detonations (or via direct carbon ignition; @pakm10 [@pakm11; @pakm12b; @kash15a]). Furthermore, prompt detonations in merging double WD binaries also have the capacity to explain the evolution of the SN Ia luminosity function. provided a baseline for radiative transfer simulations of double detonation SNe Ia by calculating the explosion and appearance of a bare C/O WD core with no overlying helium shell. They found reasonable agreement with observations of SNe Ia, both in terms of light curves and spectra. However, recent work by included a set of hydrodynamical explosions of bare C/O WDs that disagreed with the nucleosynthetic results of. Moreover, recent observational results concerning neutron-rich isotopes in SNe Ia (e.g., @seit13a [@yama15b; @dimi17a]) have been claimed as evidence against sub-\(M_{\rm Ch}\) explosions, but comprehensive in-depth studies of nucleosynthetic abundances in sub-\(M_{\rm Ch}\) detonations do not yet exist in the literature for comparison to these observations. Motivated by the disagreement in previous work and by the need for detailed nucleosynthetic data, we revisit simplified simulations of spherically symmetric bare C/O WD detonations in this paper. While recent studies have performed hydrodynamical and radiative transfer simulations with multi-dimensional helium shell ignitions (e.g., @fhr07 [@fink10; @krom10; @sim12; @moll13a]), their use of relatively massive helium shells yielded significant amounts of iron-group elements in the helium detonation ashes, which continues to be a vexing issue for obtaining spectroscopically normal SNe Ia from these models. The much smaller helium shells at ignition found by suggest that the study of one-dimensional baseline bare WD core detonations with no helium shell is still informative. Future work will continue the development of double detonation models by including these very low mass helium shells in multi-dimensional simulations. We begin in §[2](#sec:setup){reference-type="ref" reference="sec:setup"} by describing our method for artificially broadening detonations in WDs into structures that are spatially resolved on our numerical grid. In §[3](#sec:results){reference-type="ref" reference="sec:results"}, we detail our nucleosynthetic results for a suite of 80 post-processed simulations, focusing on bulk yields in §[3.1](#sec:bulk){reference-type="ref" reference="sec:bulk"} and on neutron-rich nucleosynthesis in §[3.2](#sec:nrich){reference-type="ref" reference="sec:nrich"}. We perform radiative transfer simulations and demonstrate qualitative agreement with light curves (§[4.1](#sec:lcs){reference-type="ref" reference="sec:lcs"}) and spectra (§[4.2](#sec:spectra){reference-type="ref" reference="sec:spectra"}) of observed SNe Ia, and we conclude with avenues for future research in §[5](#sec:conc){reference-type="ref" reference="sec:conc"}.
# Simulation Details {#sec:setup}
In this section, we describe our procedure for setting up, running, and post-processing our reactive hydrodynamic simulations. We begin by calculating the initial conditions for our white dwarfs (WDs) with the stellar evolution code `MESA`[^1]. We construct WDs with masses of \(0.8\), \(0.85\), \(0.9\), \(1.0\), and \(1.1 \, M_\sun\) and uniform compositions of \(50/50\) or \(30/70\) C/O by mass. The WDs are initially hot and are allowed to cool until their central temperatures reach \(\unit[3\E{7}]{K}\). The density profiles of these 10 models are then used as initial conditions for our `FLASH`[^2] simulations. `FLASH` and `MESA` use the same equation of state for most of the relevant parameter space, but there is still a small deviation from hydrostatic balance in the outer regions of the WD after mapping to `FLASH`. However, any spurious velocities are erased after the detonation passes. Each one-dimensional spherically symmetric simulation has a domain size of \(\unit[10^{11}]{cm}\) and 19 levels of adaptive mesh refinement for a minimum cell size of \(\unit[4.8\E{4}]{cm}\) within the WD. The criteria for refinement are based on the gradients of pressure, density, and temperature using `FLASH`'s default thresholds. At a radius initially just outside the WD's surface, the minimum allowed cell size increases by a factor of two and continues to increase linearly with radius beyond this location. This limits the amount of computational time spent following the shock that propagates outwards into the ambient medium after the detonation passes through the WD. Additionally, the maximum level of refinement in the innermost \(\unit[10^7]{cm}\) is reduced by four levels so that inwardly propagating shocks do not limit the global timestep as they converge towards the center and increase their velocity. The C/O ratio of the WD in `FLASH` is set to match the `MESA` model from which it came. Furthermore, we include four different metallicities for our initial models: \(0\), \(0.5\), \(1\), and \(2 \, Z_\odot\), which we approximate by including \(^{22}\)Ne, the stopping point for CNO isotopes following helium-burning, and \(^{56}\)Fe at mass fractions of \(X_{\rm 22Ne}=0\), \(0.005\), \(0.01\), and \(0.02\) with \(X_{\rm 56Fe}=0.1 \, X_{\rm 22Ne}\). The ambient medium surrounding the WD is initialized with a density and temperature of \(\unit[10^{-3}]{g \, cm^{-3}}\) and \(\unit[10^6]{K}\). We enable monopole gravity and nuclear burning. Burning in physical detonations occurs behind the nearly infinitesimally thin shock front, so reactions in `FLASH` are disabled within shocks by default to avoid unphysical detonation structures. showed that an Eulerian piecewise parabolic method hydrodynamics code with reactions disabled within shocks produces the correct detonation speeds even for unresolved burning, as well as avoiding a potential artificial deflagration caused by numerical mixing at the shock front. See Appendix A of for a more detailed discussion. We have extended `FLASH`'s nuclear burning capabilities by incorporating an interface to `MESA`'s nuclear burning module, which enables the ability to construct an arbitrary nuclear reaction network. For our hydrodynamic simulations, we use a 41-isotope network comprised of neutrons, \(^1\)H, \(^4\)He, \(^{11}\)B, \(^{12}\)C, \(^{13-14}\)N, \(^{16-17}\)O, \(^{20,22}\)Ne, \(^{23}\)Na, \(^{24-26}\)Mg, \(^{27}\)Al, \(^{28-30}\)Si, \(^{30-31}\)P, \(^{31-32}\)S, \(^{35}\)Cl, \(^{36-39}\)Ar, \(^{39}\)K, \(^{40}\)Ca, \(^{43}\)Sc, \(^{44}\)Ti, \(^{47}\)V, \(^{48}\)Cr, \(^{51}\)Mn, \(^{52,56}\)Fe, \(^{55}\)Co, and \(^{56,58-59}\)Ni, with 190 interlinking reactions from JINA's `REACLIB`. For the relevant detonation conditions in C/O-rich material, this network yields errors of at most a few percent in the energy release. Note that the above network is only tailored to track accurate energy release but not accurate isotopic abundances. In order to more precisely calculate abundances, we include tracer particles for post-processing, which track the radius, velocity, density, and temperature and are evenly spaced every \(\unit[5\E{6}]{cm}\) throughout the WD. The detonation is ignited at the center of the WD by initializing a hotspot of radius \(\unit[4\E{7}]{cm}\) that has a linear temperature gradient with a central temperature of \(\unit[2\E{9}]{K}\) and an outer temperature of \(\unit[1.2\E{9}]{K}\). The temperature just outside the hotspot and throughout the rest of the WD is set to a constant \(\unit[3\E{7}]{K}\); note that the value of the initial WD temperature is unimportant because post-shock temperatures are \(\sim 100\) times higher. The \(\unit[4\E{7}]{cm}\) hotspot is much larger than the minimum detonatable regions found by previous work but is necessary due to the burning limiter we describe below. We have confirmed that our results are insensitive to the size of the hotspot, which is reasonable because inaccurate nucleosynthesis due to the temperature perturbation will be confined to the hotspot, which corresponds to a central mass of just \(0.0014\), \(0.0018\), \(0.0025\), \(0.0045\), and \(0.0088 \, M_\sun\) for our \(0.8\), \(0.85\), \(0.9\), \(1.0\), and \(1.1 \, M_\sun\) \(50/50\) C/O WDs. One goal of our work is to ensure that we are capturing the relevant physics by spatially resolving the reaction front structure in our simulations. However, C/O detonations have lengthscales \(\unit[10-10^4]{cm}\) at our densities of interest and have thus been previously followed with a level set or progress variable method (e.g., @cald07a [@sim10; @seit13b; @town16a]). We overcome this obstacle by artificially broadening the detonation, similar in spirit to previous studies that thicken deflagration fronts, and subsequently testing our resolved simulations for convergence. We broaden the detonation by introducing a limit to the relative amount the temperature can change within each cell in one timestep due to nuclear burning, \(| \Delta \ln T |_{\rm max}\), similar to the method employed by. For our primary simulations we choose \(| \Delta \ln T |_{\rm max} = 0.04\), motivated by the convergence studies detailed in §[2.1](#sec:conv){reference-type="ref" reference="sec:conv"}. Reactions and hydrodynamics in `FLASH` are computed in an operator-split fashion. Between each computation of the hydrodynamic evolution during a time \(\Delta t_{\rm hydro}\), the timestep determined by the Courant condition, a temporally resolved integration of the reactions is performed in each cell with an initial integration time of \(\Delta t_{\rm react}=\Delta t_{\rm hydro}\). From the entropy equation, the relative change in temperature is \(\Delta \ln T \sim \bar{ \epsilon } \Delta t_{\rm react} / c_V T\), where \(\bar{ \epsilon }\) is the average energy generation rate over the timestep \(\Delta t_{\rm react}\), and \(c_V\) is the specific heat at constant volume. If \(| \Delta \ln T | > | \Delta \ln T |_{\rm max}\) in a cell, the burning integration time, \(\Delta t_{\rm react}\), is reduced to the appropriate value via a Newton-Raphson iteration while leaving \(\Delta t_{\rm hydro}\) unchanged. The burning integration is re-run for each iteration in order to yield consistent energetics and abundance changes. This limiting procedure can also be thought of as integrating the reactions for the full hydrodynamic timestep, \(\Delta t_{\rm hydro}\), but with all the reaction rates reduced by the multiplicative factor \(\Delta t_{\rm react} / \Delta t_{\rm hydro}\). Simulations are evolved for \(\unit[10]{s}\), after which the tracer particles' density and temperature histories are post-processed with `MESA`'s one zone burner. We employ a \(205\)-isotope network that includes neutrons, \(^{1-2}\)H, \(^{3-4}\)He, \(^{6-7}\)Li, \(^{7,9-10}\)Be, \(^{8,10-11}\)B, \(^{12-13}\)C, \(^{13-16}\)N, \(^{15-19}\)O, \(^{17-20}\)F, \(^{19-23}\)Ne, \(^{21-24}\)Na, \(^{23-27}\)Mg, \(^{25-28}\)Al, \(^{27-33}\)Si, \(^{30-34}\)P, \(^{31-37}\)S, \(^{35-38}\)Cl, \(^{35-41}\)Ar, \(^{39-44}\)K, \(^{39-49}\)Ca, \(^{43-51}\)Sc, \(^{43-54}\)Ti, \(^{47-56}\)V, \(^{47-58}\)Cr, \(^{51-56}\)Mn, \(^{51-62}\)Fe, \(^{54-62}\)Co, \(^{54-62}\)Ni, \(^{58-66}\)Cu, \(^{59-66}\)Zn, \(^{59-66}\)Ga, and \(^{59-66}\)Ge and interlinking reactions from JINA's `REACLIB`. We post-process each of our 40 hydrodynamic simulations with two different normalizations of the \(^{12}\)C\(+^{16}\)O reaction rate (\(1\) and \(0.1\) times the default rate; see §[3.2.4](#sec:SNR_CrFeCaS){reference-type="ref" reference="sec:SNR_CrFeCaS"} for the motivation behind this variation) for a total of 80 post-processed results.
## Convergence studies {#sec:conv}
In this section, we demonstrate the convergence of our results as we increase the resolution in our simulations for a set of \(1.0 \, M_\sun\) \(50/50\) C/O solar metallicity WD detonations. Note that since the physical burning scales are not resolved by many orders of magnitude, convergence does not imply correctness, only that our thickening scheme is numerically consistent over the range of grid scales used here. Verification of yields against resolved calculations will be the topic of future work. Figure [\[fig:resolution\]](#fig:resolution){reference-type="ref" reference="fig:resolution"} shows synthesized masses of \(^{56}\)Ni and intermediate-mass elements (IMEs; defined as having charges \(11 \leq Z \leq 20\)) and the total nuclear energy release, \(E_{\rm nuc}\), for three sets of hydrodynamic simulations and three post-processed results vs. the minimum cell size in the simulation. The dashed lines represent the hydrodynamic results, which use a 41-isotope network, and the solid lines show results from post-processing the same hydrodynamic simulations using a \(205\)-isotope network. As Figure [\[fig:resolution\]](#fig:resolution){reference-type="ref" reference="fig:resolution"} demonstrates, global values are converged for minimum cell sizes \(\lesssim \unit[10^5]{cm}\) for both the \(| \Delta \ln T |_{\rm max}=0.02\) and \(0.04\) cases, with relevant quantities changing by \(< 1\%\) with a factor of two increase in resolution. Results for the \(| \Delta \ln T |_{\rm max}=0.08\) simulation do not appear to be fully converged at our highest resolution, which motivates our choice of \(| \Delta \ln T |_{\rm max}=0.04\) for all of the production runs in this work. The bulk nucleosynthetic yields of the hydrodynamic results without post-processing and the results after post-processing are discrepant at a \(3-10\%\) level. However, as previously discussed, the 41-isotope nuclear reaction network used in the hydrodynamics simulations is designed to capture energetics, not isotopic abundances. Thus, the agreement in energetics before and after post-processing is much better, with only a \(\simeq 0.3\%\) difference. This gives confidence that the tracer particles' density and temperature histories used in the post-processing calculation and the resulting nucleosynthetic abundances are accurate.
## Spatially resolved broadened detonation structure {#sec:resolved}
Our burning limiter allows us to spatially resolve the artificially broadened detonation structure in our hydrodynamic simulations, an example of which is shown in Figure [\[fig:lxvslr\]](#fig:lxvslr){reference-type="ref" reference="fig:lxvslr"}. The top panel shows thermodynamic variable profiles, and the bottom panel shows profiles of the energy generation rate normalized to the maximum value, \(\epsilon / \epsilon_{\rm max}\), and the mass fractions of 11 isotopes as labeled. The other 30 isotopes comprising the 41-isotope network used in our hydrodynamical simulations do not reach mass fractions above \(10^{-2}\) in this plot at this time, \(\unit[0.24]{s}\) after the simulation has begun. The time of the snapshot is chosen to coincide with when the detonation reaches the mass coordinate (\(0.64 \, M_\sun\) from the center) where the \(^{56}\)Ni fraction will equal the \(^{28}\)Si fraction after the simulation ends. The density upstream of the detonation at this time is \(\unit[6.3\E{6}]{g \, cm^{-3}}\). The carbon consumption lengthscale for a steady state detonation at this density is \(\sim \unit[10^2]{cm}\), and the lengthscale for an overdriven detonation such as this is even shorter. Due to the use of our burning limiter, we achieve a spatially resolved detonation by construction. The broadened detonation in our simulation has a carbon consumption lengthscale of \(\sim \unit[10^5]{cm}\), \(>10\) times longer than the true lengthscale, and the maximum of the energy generation rate is several zones behind the shock front instead of just behind or inside the shock where it would be located for an unresolved detonation. While the detonation itself is not physically correct, the convergence study in §[2.1](#sec:conv){reference-type="ref" reference="sec:conv"} gives us confidence that the major yields will be relatively unchanged at higher resolutions. These yields will be verified by comparison to resolved calculations in future work.
## Ejecta profiles {#sec:rhovsv}
Figure [\[fig:rhovsv\]](#fig:rhovsv){reference-type="ref" reference="fig:rhovsv"} shows density vs. velocity profiles \(\unit[10]{s}\) after the simulation begins for our five WD masses with an initial C/O mass fraction of \(50/50\). The profiles are all relatively similar: a nearly constant density core surrounded by an exponentially declining density beyond \(\sim \unit[10^4]{km \, s^{-1}}\). Studies of SN Ia ejecta interaction with surrounding material often use an exponential parameterized approximation to the ejecta profile (e.g., @dwar98a): \[\rho(v,t) = \frac{6^{3/2}}{8 \pi} \frac{ M_e^{5/2}}{E_{\rm kin}^{3/2}} \frac{ \exp{ \left(-v/v_e \right) } }{t^3},\] where the kinetic energy is \(E_{\rm kin}\), the ejecta mass is \(M_e\), and \(v_e = \left( E_{\rm kin} / 6 M_e \right)^{1/2}\). We plot this parameterization for our \(1.0 \, M_\sun\) model as a yellow dotted line. We also plot the exponential parameterization of 's \(M_{\rm Ch}\) W7 model as a dashed line for comparison. In the outer regions \(\geq \unit[10^4]{km \, s^{-1}}\), the exponential approximation provides a reasonable fit to our model. However, in the inner \(0.2 \, M_\sun\), the exponential parameterization of our model and of W7 yield substantially higher densities with a steeper slope than found in our simulations. These differences will have a significant impact on modeling of the nebular and SN remnant phases, when these inner regions become optically thin. Indeed, have recently found better agreement with the nebular spectra of SN 2011fe when using parameterized ejecta profiles with constant density cores instead of exponential profiles. Future modeling of nebular spectra and emission from SN remnants using the ejecta profiles from our hydrodynamic simulations will enable more quantitative comparisons to observations.
# Nucleosynthetic results {#sec:results}
We now describe the nucleosynthetic products of our post-processed models. After presenting the bulk yields and comparing them to previous work, we will discuss our trace abundances in the context of observations from late-time SN Ia light curves, the solar Mn abundance, and SN remnant abundances. These observations constrain the amount of neutron-rich nucleosynthesis in SNe Ia, an important discriminant between \(M_{\rm Ch}\) and sub-\(M_{\rm Ch}\) progenitors.
## Bulk yields and comparison to literature {#sec:bulk}
In this section, we report the yields of low-mass elements (LMEs; \(Z \le 10\)), IMEs, high-mass elements (HMEs; \(21 \le Z\)), and \(^{56}\)Ni and compare our results to previous work.
### Yield profiles and integrated masses
In Figure [\[fig:xvsm\]](#fig:xvsm){reference-type="ref" reference="fig:xvsm"}, we show mass fractions of LMEs, IMEs, HMEs excluding \(^{56}\)Ni, and \(^{56}\)Ni vs. mass coordinate. The five panels represent the post-processing results of different WD masses (\(0.8-1.1 \, M_\sun\) from top to bottom) with initial compositions of \(50/50\) C/O and solar metallicity. Also marked are the mass coordinates of velocities in increments of \(\unit[5000]{km \, s^{-1}}\). The profiles show stratified composition structures as expected for one-dimensional pure detonations with no mixing. \(^{56}\)Ni and other HMEs are produced in the center of the WDs and extend out to varying mass coordinates depending on the WD mass. This material is surrounded by a layer of IMEs, which is in turn surrounded by a LME cap primarily composed of \(^{16}\)O. One interesting feature is the presence of \(^4\)He with a mass fraction of \(\sim 0.01\) in the central few tenths of a solar mass of the more massive \(1.0\) and \(1.1 \, M_\sun\) WDs. This is indicative of the \(\alpha\)-rich freezeout from nuclear statistical equilibrium (NSE) characteristic of nuclear burning at these temperatures and densities, which will have an effect on the production of neutron-rich isotopes discussed in §[3.2](#sec:nrich){reference-type="ref" reference="sec:nrich"}. The presence of \(^4\)He in the core, mixed with \(^{56}\)Ni, could result in an interesting signature in late-time nebular spectra; we leave an analysis of its effect to future work. Figure [\[fig:mvsm\]](#fig:mvsm){reference-type="ref" reference="fig:mvsm"} shows post-processed results for total synthesized masses vs. WD mass for an initial C/O mass fraction of \(50/50\) and four initial metallicities. Increasing the metallicity increases the non-\(^{56}\)Ni HME mass but decreases the \(^{56}\)Ni mass; the LME and IME masses are relatively constant with respect to the metallicity. The \(^{56}\)Ni dependence on the metallicity for our high-mass models is similar to that found for \(M_{\rm Ch}\) explosions. We obtain a \(\sim 10\%\) decrease in \(^{56}\)Ni mass for a \(1.0 \, M_\sun\) WD detonation when the initial metallicity is changed from \(0\) to \(2 \, Z_\odot\). However, there is a more drastic dependence for the low-mass models: a zero metallicity \(0.8 \, M_\sun\) WD detonation produces almost a factor of two more \(^{56}\)Ni than a \(2 \, Z_\odot\) explosion.
### Comparison of bulk yields to other results {#sec:bulkcomp}
In Figure [\[fig:mvsm_comp\]](#fig:mvsm_comp){reference-type="ref" reference="fig:mvsm_comp"}, we show a comparison of our hydrodynamic and post-processed bulk yields to previous work. The top, middle, and bottom panels show the ratios of total synthesized masses of \(^{56}\)Ni, IMEs, and \(^{16}\)O, respectively, to our post-processed results. Our hydrodynamic results for an initial composition of \(50/50\) C/O and zero metallicity are shown as thin lines, and the post-processed results are shown as thick lines. Green circles represent zero metallicity synthesized masses from, yellow triangles demarcate 's \(^{56}\)Ni masses with an initial metallicity of \(\sim 2 \, Z_\odot\), blue crosses and plus signs are zero metallicity \(^{56}\)Ni masses resulting from 19-isotope and 199-isotope simulations by, and magenta diamonds are \(\sim Z_\odot\) results from, respectively. The ratios are calculated using our post-processed yields from models with the appropriate initial metallicity. Red stars represent zero metallicity results from a parameterized model for burning in `FLASH` in which the detonation front is tracked by progress variables that measure the fractions of fuel, ash, quasi-NSE material, and NSE material. This front tracking scheme is used in a hydrodynamic `FLASH` simulation with minimum cell size of \(\unit[1.25\E{4}]{cm}\) and zero metallicity, whose results are then post-processed with the same \(205\)-isotope network used throughout the rest of this work. A similar procedure was also used in. The results of our hydrodynamic and post-processed burning limiter simulations are very similar to the parameterized model results using progress variables, which has been verified against resolved calculations of planar steady-state detonations, giving us further confidence that our results are converged. The burning in both methods is systematically more complete (e.g., more \(^{56}\)Ni is produced) than all of the other studies except for the large network results of at low WD masses \(\leq 0.9 \, M_\sun\). For a WD mass of \(0.9\) \((1.0) \, M_\sun\), our post-processed model yields a \(^{56}\)Ni mass of \(0.30\) \((0.58) \, M_\sun\), while a quadratic fit to 's results implies a mass of \(0.11\) \((0.38) \, M_\sun\). These abundance differences will be reflected in our radiative transfer predictions (§[4](#sec:rad){reference-type="ref" reference="sec:rad"}), enabling typical SNe Ia to be produced by \(1.0 \, M_\sun\) WDs instead of \(1.1 \, M_\sun\) WDs as found by. This will imply, among other things, a higher predicted rate of SNe Ia because less massive WDs are more numerous. It is also apparent that the total mass burned in 's simulations is more steeply dependent on WD mass than we have found. This likely contributes to the difference in the slope of the brightness-decline rate relation that we show in §[4](#sec:rad){reference-type="ref" reference="sec:rad"}. It is unclear why 's nucleosynthetic results differ so significantly from ours. We note that 's \(^{56}\)Ni masses are in rough agreement with those of in their limited mass range (yellow triangles in the top panel of Fig. [\[fig:mvsm_comp\]](#fig:mvsm_comp){reference-type="ref" reference="fig:mvsm_comp"}), especially after adjusting for 's initial composition of zero metallicity and 's \(\sim 2 \, Z_\odot\) initial composition. However, possibly due to a neglect of Coulomb corrections, the central densities reported by are systematically lower than we,, and others calculate, and thus their derived \(^{56}\)Ni masses will also be lower. Therefore, 's agreement with is consistent with both of their reported \(^{56}\)Ni masses being too low. The discrepancy between our results and those of, and to a lesser extent the \(19\)-isotope calculations of, is easier to explain. Smaller networks may neglect burning pathways that become increasingly important for lower density, low mass WDs. This is particularly true for the 4-stage network used by. The discrepancy is less severe for higher WD masses because much of the IGE nucleosynthesis occurs in NSE, which erases details of the nuclear reaction network and the detonation structure. However, their \(0.88 \, M_\sun\) model produces just \(1/3\) of the \(^{56}\)Ni that our calculations imply. Such a large difference in \(^{56}\)Ni abundance will have a significant impact on radiative transfer calculations, particularly for subluminous SNe Ia, an effect we will discuss in more detail in §[4](#sec:rad){reference-type="ref" reference="sec:rad"}.
## Neutron-rich nucleosynthesis {#sec:nrich}
While simulations of deflagrations, detonations, and deflagration-to-detonation transitions of C/O WDs generally produce similar bulk nucleosynthetic results at the order of magnitude level, the different explosion mechanisms yield large differences in trace abundances. This is especially true for neutron-rich isotopes. The higher densities and longer timescales involved in \(M_{\rm Ch}\) deflagration-to-detonation transition explosions allow for weak reactions that can significantly reduce the electron fraction from its initial value close to \(0.5\). Some neutron-rich isotopes are produced in our pure detonation simulations, particularly in regions that undergo incomplete silicon-burning, but the overall abundances are lower due to the \(\alpha\)-rich freezeout from NSE that occurs in the core. Some models of nebular spectra have implied the production of up to \(0.2 \, M_\sun\) of neutron-rich stable IGEs in the center of SN Ia ejecta (e.g., @mazz07 [@mazz15a]). However, there is some disagreement about the required amount of stable IGEs, in part due to uncertainties in the ejecta density profile (§[2.3](#sec:rhovsv){reference-type="ref" reference="sec:rhovsv"}). found that the sub-\(M_{\rm Ch}\) double detonation model of with \(0.02 \, M_\sun\) of stable IGEs provides the best density and composition profile for a nebular spectrum of SN 1994D. More recently, arrive at the conclusion that a stable IGE core is not required to match the nebular spectra of SN 2011fe and may in fact be disfavored. Deriving the amount of stable IGE from nebular spectra is complicated by the fact that if there is a surviving WD companion, it will capture some \(^{56}\)Ni from the SN ejecta. Some of this accreted \(^{56}\)Ni will be hot enough to be fully ionized and will have a slower rate of decay due to its inability to capture electrons. Thus, there may be an additional source of heating that is currently unaccounted for in nebular phase studies, which will change the masses inferred from observations. We leave a detailed study of the nebular spectra expected from our pure detonation models to future work. In the following sections, we explore other probes of neutron-rich nucleosynthesis: late-time light curve observations, the solar abundance of Mn, and abundance estimates from SN remnant observations.
### Late-time light curve observations {#sec:latetime}
Several of the neutron-rich isotopes produced in SNe Ia have a significant impact on the late-time light curves after \(\unit[800]{d}\). At these late phases, \(\gamma\)-ray trapping is inefficient, and the predominant energy source is the thermalization of positron and electron kinetic energy. These leptons arise from the decay of \(^{56}\)Co (half-life of \(\unit[77]{d}\), produced primarily as \(^{56}\)Ni) and the neutron-rich isotopes \(^{57}\)Co (half-life of \(\unit[272]{d}\), produced primarily as \(^{57}\)Ni) and \(^{55}\)Fe (half-life of \(\unit[1000]{d}\), produced primarily as \(^{55}\)Co). Several recent nearby SNe Ia (SN 2011fe, SN 2012cg, and SN 2014J) have been observed to late enough phases to estimate the abundances of these neutron-rich isotopes from their contribution to the light curve. The implied mass ratio of \(^{57}\)Co to \(^{56}\)Co at these late times ranges from \(0.02\) to \(0.09\), while the \(^{55}\)Fe to \(^{57}\)Co mass ratio has been estimated to be \(<0.2\), albeit with large error bars. In Figure [\[fig:nrich57\]](#fig:nrich57){reference-type="ref" reference="fig:nrich57"}, we show the mass ratios of \(^{57}\)Ni to \(^{56}\)Ni and \(^{55}\)Co to \(^{57}\)Ni produced in our explosions for a range of metallicities. The initial C/O ratio for all models is \(50/50\). Changing the initial C/O ratio alters the mass ratios at a minimal level; we do not plot these results for simplicity. The upper limit to the \(^{55}\)Co to \(^{57}\)Ni mass ratio from is shown as a blue dotted line, and the range of \(^{57}\)Ni to \(^{56}\)Ni ratios inferred from observations is shown as a red shaded region. The increase in the \(^{57}\)Ni/\(^{56}\)Ni ratio with mass for masses \(\geq 0.9 \, M_\sun\) is due to the changing detonation regimes: as the WD mass increases, the primary mode of burning transitions from incomplete silicon-burning to an \(\alpha\)-rich freezeout from NSE, with an accompanying change in the \(^{57}\)Ni/\(^{56}\)Ni ratio. However, the reason for the decrease in the ratio with increasing mass below \(0.9 \, M_\sun\) is uncertain. Similarly, the origin of the large gap in the \(^{55}\)Co/\(^{57}\)Ni ratio between zero and half solar metallicity models is unknown. This gap is driven by the metallicity dependence of the \(^{55}\)Co yield, which is also displayed in Figure [\[fig:nrich\]](#fig:nrich){reference-type="ref" reference="fig:nrich"}, but the reason for this dependence is unclear. We leave exploration of these trends to future work. Our \(1.1 \, M_\sun\) results agree broadly with 's values for a \(0.9+1.1 \, M_\sun\) violent merger of two WDs, whose nucleosynthesis is primarily determined by the explosion of the more massive WD. Our results for the range of masses and metallicities do not alter the tension between the low \(^{57}\)Ni masses produced in sub-\(M_{\rm Ch}\) detonation models and the higher masses inferred from late-time observations. However, the \(^{57}\)Ni and \(^{55}\)Co masses derived from observations have very large error bars due to the possible contribution of light echoes and uncertainties in the \(\gamma\)-ray and lepton trapping efficiencies. Furthermore, the possibility of a surviving companion WD that complicates nebular spectra modeling will also have an influence here. If a companion WD survives the SN Ia explosion, it will capture a small amount of \(^{56}\)Ni. The radioactive decay of this accreted ejecta will be delayed due to the fully ionized \(^{56}\)Ni's inability to capture electrons, and so the surviving companion WD can supplement the SN Ia ejecta's late-time luminosity. This additional luminosity will reduce the amount of \(^{57}\)Co inferred from observations and possibly bring our nucleosynthetic results into agreement. Ongoing and future late-time observations, particularly of SN 2011fe and SN 2014J, will shed further light on this issue; for now, we do not regard this tension as strong evidence against sub-\(M_{\rm Ch}\) detonation models.
### Solar abundance of manganese {#sec:solarMn}
The production, and subsequent decay, of the neutron-rich isotope \(^{55}\)Fe in SNe Ia contributes to the late-time luminosity, as described in the previous section, and is also the primary source of \(^{55}\)Mn in the Sun. argue that the known non-SN Ia sites of nucleosynthesis produce a sub-solar ratio of Mn to Fe after all relevant radioactive decays have occurred, and thus SNe Ia must make up the difference by producing a super-solar Mn/Fe ratio. Because their representative sub-\(M_{\rm Ch}\) model (a \(0.9+1.1 \, M_\sun\) violent merger of two WDs; @pakm12b) has a sub-solar Mn/Fe ratio, conclude that \(\sim 50\%\) of SNe Ia must occur via a deflagration-to-detonation transition explosion in a \(M_{\rm Ch}\) WD. In Figure [\[fig:nrich\]](#fig:nrich){reference-type="ref" reference="fig:nrich"}, we show the \(^{55}\)Mn mass produced in our post-processed simulations, after accounting for all radioactive decays, vs. WD mass for an initial C/O mass ratio of \(50/50\). Red lines show the mass ratio of Mn to Fe, again after all decays have occurred. As before, changing the initial C/O ratio has a minimal effect on these results, so these models are omitted for simplicity. For our lowest mass models, \(^{55}\)Co, which eventually decays to \(^{55}\)Mn, is produced via incomplete silicon-burning. As the WD mass and central density are increased, more of the WD core undergoes incomplete silicon-burning, so the final \(^{55}\)Mn yield increases. However, the \(^{56}\)Ni yield increases more strongly with WD mass, so the overall final Mn/Fe ratio decreases. As the WD mass increases past \(\sim 0.9 \, M_\sun\), some of the detonated material enters the regime of \(\alpha\)-rich freezeout from NSE, which reduces the yield of \(^{55}\)Co and the Mn/Fe ratio. Presumably, for even higher mass pure detonation models, the detonated material will reach conditions for a "normal" freezeout from NSE, and the \(^{55}\)Co yield will again increase with mass, but our highest mass WD explosions are not yet in this regime. The solar value of the Mn/Fe mass ratio is shown as a red dotted line. In agreement with 's \(1.1 \, M_\sun\) WD detonation, our higher-mass \(\ge 1.0 \, M_\sun\) models yield sub-solar Mn/Fe mass ratios. However, a super-solar value is achieved for lower-mass \(\le 0.9 \, M_\sun\) detonations at an initial metallicity of \(0.5 \, Z_\odot\). Thus, at least part of the discrepancy found by between the solar Mn/Fe ratio and nucleosynthesis in sub-\(M_{\rm Ch}\) detonations can be alleviated by including pure detonations of lower-mass WDs. However, unless lower-mass detonations significantly outnumber higher-mass explosions, it is not clear that only including core collapse SNe and sub-\(M_{\rm Ch}\) WD detonations will yield the solar Mn/Fe value. The possibility remains that a combination of core collapse SNe, sub-\(M_{\rm Ch}\) WD detonations, and the class of peculiar Type Iax SNe may yield the correct solar value, or that \(M_{\rm Ch}\) explosions do indeed contribute to Mn production but at a lower fraction of all SNe Ia; further work is still required to solve this issue.
### SN remnant observations: Mn/Fe vs. Ni/Fe {#sec:SNR_MnFeNiFe}
SN remnants serve as another probe of detailed nucleosynthesis in SN Ia explosions. As the ejecta sweeps up the surrounding interstellar medium, a reverse shock propagates into the ejecta, exciting it to X-ray-emitting temperatures. The resulting emission can be used to infer nucleosynthetic yields, although the process is complicated by noisy spectra, non-equilibrium ionization effects, asymmetric and inhomogeneous density distributions, and incomplete propagation of the reverse shock into the ejecta. One such probe of SN Ia combustion conditions, the mass ratios of Mn/Fe and Ni/Fe, was examined by. In Figure [\[fig:mnfenife\]](#fig:mnfenife){reference-type="ref" reference="fig:mnfenife"}, we compare our post-processed nucleosynthetic mass ratios of Mn/Fe vs. Ni/Fe to their observational results, shown as gray symbols. Five WD masses at four initial metallicities are shown for an initial C/O mass fraction of \(50/50\). Our ratios are calculated during the SN remnant phase, which for practical purposes we take to be between \(\unit[10^2-10^5]{yr}\). We thus account for isotopes that are present during this phase but ultimately decay to another element. For example, the Mn present during the SN remnant phase is predominantly the stable isotope \(^{55}\)Mn, but there is a small contribution from \(^{53}\)Mn, which decays to \(^{53}\)Cr with a half-life of \(\unit[4\E{6}]{yr}\). Thus, the Mn masses in Figures [\[fig:nrich\]](#fig:nrich){reference-type="ref" reference="fig:nrich"} and [\[fig:mnfenife\]](#fig:mnfenife){reference-type="ref" reference="fig:mnfenife"} differ slightly. Likewise, the Ni present during the SN remnant phase is dominated by the stable isotopes \(^{58}\)Ni, \(^{60}\)Ni, and \(^{62}\)Ni, but the isotope \(^{59}\)Ni, with a half-life of \(\unit[8\E{4}]{yr}\) can contribute a few percent by mass. Our results are consistent with the sub-\(M_{\rm Ch}\) detonation results calculated in. Thus, we also agree that matching the Tycho and Kepler SN remnant compositions requires somewhat super-solar metallicities, and that the composition of 3C 397 implies an unrealistically high initial metallicity if it was the product of a sub-\(M_{\rm Ch}\) explosion. claim that this mismatch is evidence for a \(M_{\rm Ch}\) explosion, but we emphasize that a \(M_{\rm Ch}\) explanation also requires an extremely high metallicity, a complicated ejecta geometry, an unexpectedly high central density, or a combination of all three. Thus, the abundances in SN remnant 3C 397 continue to present a nucleosynthetic puzzle for any standard scenario. The implication that Tycho and Kepler's exploding WDs had super-solar metallicities is also somewhat problematic, given the solar or slightly sub-solar metallicities of the stellar environments at their Galactocentric radii. However, this discrepancy can be at least partially explained by the fact that these remnants are young and their reverse shocks have not fully traversed the SN ejecta. Thus, the inferred mass ratios may not be representative of the ejecta's total nucleosynthesis. In Figure [\[fig:mnfenife_frac\]](#fig:mnfenife_frac){reference-type="ref" reference="fig:mnfenife_frac"}, we show Mn/Fe vs. Ni/Fe mass ratios for our \(0.9 \, M_\sun\) models for a range of reverse-shocked ejecta fractions. 3C 397's SN remnant is likely fully reverse-shocked, so this analysis does not apply to it. However, Figure [\[fig:mnfenife_frac\]](#fig:mnfenife_frac){reference-type="ref" reference="fig:mnfenife_frac"} shows that Tycho and Kepler's SNe may be explained as the explosions of \(\sim 0.25 \, Z_\odot\) sub-\(M_{\rm Ch}\) WDs with young remnants whose reverse shocks have only encountered \(25-50\%\) of the total ejecta. Given the Galactic positions of the SNe, the implication of highly sub-solar metallicity progenitors is not any more reasonable than the super-solar metallicities inferred from Figure [\[fig:mnfenife\]](#fig:mnfenife){reference-type="ref" reference="fig:mnfenife"}. However, this analysis demonstrates the difficulty of ruling out progenitor models for young SN remnants using this particular diagnostic.
### SN remnant observations: Cr/Fe vs. Ca/S {#sec:SNR_CrFeCaS}
We now turn to an exploration of Cr/Fe vs. Ca/S mass ratios during the SN remnant phase, motivated by the work of. While none of these isotopes directly traces neutron-rich nucleosynthesis (Cr and Fe are primarily produced as \(^{52}\)Fe and \(^{56}\)Ni, respectively, during the explosion, which have equal numbers of protons and neutrons), the Ca/S ratio does have an inverse correlation with the neutron excess at the time of explosion. In Figure [\[fig:CrFeCaS\]](#fig:CrFeCaS){reference-type="ref" reference="fig:CrFeCaS"}, we show our post-processed results for the Cr/Fe mass ratio vs. the Ca/S mass ratio during the SN remnant phase, accounting for intermediate decays as before. All 80 nucleosynthetic calculations are shown, corresponding to five WD masses, four metallicities, initial C/O mass fractions of \(50/50\) and \(30/70\), and two choices for the \(^{12}\)C\(+^{16}\)O reaction rate: the default `REACLIB` reaction rate and the rate scaled by a multiplicative factor, \(f_{\rm ^{12}C+^{16}O}=0.1\), as motivated by. Observational values for five Galactic and LMC remnants from are shown in gray. The remnant G337.2 does not have constraints on its Cr/Fe mass ratio, so it is shown as a vertical band. It is clear that only very low metallicity \(30/70\) C/O explosions in the top panel are consistent with the observed SN remnants. As previously mentioned, there is some uncertainty in the fact that some of these remnants may not be old enough to have their entire ejecta traversed by the reverse shock, so that the mass ratios inferred from observations may not be representative of the entire ejecta. However, the primary discrepancy lies in the Ca/S ratio, and since these IMEs are located in the outer parts of the ejecta, they have likely already been excited by the reverse shock. Much better agreement is found in the bottom panel, for which the \(^{12}\)C\(+^{16}\)O reaction rate is reduced by a factor of 10. Here, solar and sub-solar metallicities and C/O ratios of both \(50/50\) and \(30/70\) match values for observed SN remnants. Our results are consistent with 's findings; as they explain, a slower \(^{12}\)C\(+^{16}\)O reaction rate increases the abundance of \(^4\)He nuclei, which favors the production of isotopes higher in the \(\alpha\)-chain, and thus a higher Ca/S ratio. However, the \(^{12}\)C\(+^{16}\)O is not actually uncertain to a factor of 10. Unlike for the typical relatively low-energy stellar case, reaction rates at energies relevant to stellar detonations can be probed in the laboratory. The burning temperature \(\sim \unit[4\E{9}]{K}\) of the carbon detonation yields a Gamow peak of \(\unit[7.7]{MeV}\) with width \(\unit[3.8]{MeV}\), an energy range at which the cross-section of the \(^{12}\)C\(+^{16}\)O reaction has been directly measured. The *S* factor at the Gamow peak has an experimental uncertainty of only \(\sim 50\%\), and its median is actually \(\sim 20\%\) higher than the value used in `REACLIB`. The uncertainty at lower energies within the peak is higher, a factor of \(\sim 2\), but since the rate is dominated by the cross-section near the peak's maximum, the rate is only uncertain by \(\sim 50\%\) at our temperatures of interest. Thus, while we do find good agreement with the observed Ca/S ratio in the Tycho and Kepler SN remnants for near-solar metallicities and a reduced \(^{12}\)C\(+^{16}\)O reaction rate, this is not a likely explanation. Using the default `REACLIB` rate, our sub-\(M_{\rm Ch}\) models imply low metallicity progenitors for these remnants. However, we note that the \(M_{\rm Ch}\) models in yield a similar conclusion when the default \(^{12}\)C\(+^{16}\)O reaction rate is used.
# Radiative transfer calculations {#sec:rad}
A stringent test of the validity of our sub-\(M_{\rm Ch}\) WD detonation models is a comparison to the rich SN Ia observational data sets collected in the past few decades. To this end, we employ the Monte Carlo radiative transfer code SEDONA to produce synthetic light curves and spectra, which we discuss and compare to observations in the following sections. These calculations assume the level populations are in local thermodynamic equilibrium (LTE) and that lines are purely absorbing. Note that we only consider comparisons to "normal" SNe Ia, ranging from SN 1991bg-likes to SN 1991T-likes, and not to the peculiar classes of Ca-rich transients (e.g., @kasl12) and SNe Iax (e.g., @fole13a).
## Light curves {#sec:lcs}
Figures [\[fig:lcs\]](#fig:lcs){reference-type="ref" reference="fig:lcs"} and [\[fig:lcs_3070\]](#fig:lcs_3070){reference-type="ref" reference="fig:lcs_3070"} show bolometric and broad-band light curves for post-processed models with solar metallicity and initial C/O mass fractions of \(50/50\) and \(30/70\), respectively. Vega magnitudes are used here and in the following. Also overlaid for comparison in gray are three well-observed SNe Ia: the subluminous 1991bg-like SN 1999by, the normal SN 2011fe, and the over-luminous 1991T-like SN 1999dq.[^3] The general shapes of our synthetic bolometric light curves show good agreement with observed SNe Ia. The subluminous SN 1999by is reasonably well fit by our \(0.85 \, M_\sun\) \(30/70\) C/O model, the normal SN 2011fe agrees with the \(1.0 \, M_\sun\) models, and the over-luminous SN 1999dq is somewhat brighter than our \(1.1 \, M_\sun\) models. However, there are some discrepancies in the filtered light curves. In particular, our synthetic light curves generally fall too rapidly in the \(U\) and \(B\) bands and remain too bright in the \(R\) and \(I\) bands. Our results are in broad agreement with those of, although our different nucleosynthetic output precludes an exact comparison. One notable difference is obvious after \(\unit[30]{d}\), when our bluer light curves deviate from observations, whereas 's flatten and provide a better match to observations. Since this difference persists for higher WD masses, where 's and our nucleosynthetic results do not differ drastically, the discrepancy may be due to different treatments of radiative transfer. As noted by and, different radiative transfer codes produce somewhat different light curves for the same input; further work is necessary to ascertain the cause of the mismatch. The discrepancies in filtered light curves between our results and observations can also be seen in Figure [\[fig:phillips\]](#fig:phillips){reference-type="ref" reference="fig:phillips"}, which compares the peak \(B\)-band absolute magnitude to the decline in magnitudes \(\unit[15]{d}\) after maximum, \(\Delta {\rm m_{15}}(B)\). Our solar metallicity models are shown for our five WD masses and two initial C/O fractions. Gray error bars are values from the CfA light curve data set, and black symbols are the well-observed SNe Ia used in our light curve comparisons. Very promisingly, our models reproduce the basic trend of the relation, with more massive WDs yielding brighter SNe Ia that decline more slowly than SNe Ia from less massive WDs. The agreement is far from exact, though. Similarly to, our high-mass WDs \(\ge 1.0 \, M_\sun\) lie to the right of the observed relation: they evolve too rapidly compared to observed SNe. However, as compared to, our low-mass WD detonations are brighter and evolve slightly more slowly because of the increased amount of \(^{56}\)Ni and other IGEs; e.g., our \(0.9 \, M_\sun\) models are \(1.5\) magnitudes brighter and decline \(0.1\) magnitudes less after \(\unit[15]{d}\) than a \(0.9 \, M_\sun\) explosion interpolated between their \(0.88\) and \(0.97 \, M_\sun\) models. Thus, unlike for, our \(0.85\) and \(0.9 \, M_\sun\) WD models follow the faint-end slope of the Phillips relation; our \(0.85 \, M_\sun\) \(30/70\) C/O model has a similar peak \(B\)-band magnitude and \(\Delta {\rm m_{15}}(B)\) to those of SN 1991bg-like SNe. The \(B\)-band decline rate of our model light curves is highly sensitive to line blanketing effects. The fact that our \(1.0\) and \(1.1 \, M_\sun\) models predict too rapid a decline could be related to limitations in the transport calculations. In particular, the LTE assumption adopted here, which only approximates the more complex redistribution of photons to longer wavelengths due to fluorescence, may overestimate the rate of light curve reddening. As mentioned above, 's \(U\)-and \(B\)-band light curves show some late-time flattening, which ours do not. This difference may be related to their use of a method intended to mimic non-LTE effects. The importance of these effects is supported by 's non-LTE radiative transfer calculations. The light curves of their \(\geq 1.0 \, M_\sun\) models decline more slowly than ours and those of, and they are able to much more closely match the bright end of the Phillips relation. As discussed in §[3.1.2](#sec:bulkcomp){reference-type="ref" reference="sec:bulkcomp"}, the \(^{56}\)Ni yields do not differ significantly among the various studies at these relatively high masses, and thus the differences in the light curves of the high mass explosions may be attributed to their inclusion of non-LTE effects. At the low-mass end, 's non-LTE radiative transfer calculations do not reproduce the Phillips relation, instead yielding light curves that are too dim. However, our \(^{56}\)Ni yields are several times higher than theirs in this regime. It is thus possible that a combination of the nucleosynthesis from our large network, broadened detonation simulations and non-LTE radiation transport calculations will reproduce the entirety of the Phillips relation; such a study is currently underway. We note that our \(0.8 \, M_\sun\) models do not appear to match any observed SNe Ia. As argued by, this may be due to a physical minimum WD mass and associated central density that can be ignited via a converging shock: WDs that are too low in mass cannot explode as double detonation SNe Ia. Given the qualitative agreement between our \(0.85 \, M_\sun\) models and SN 1991bg-like SNe, the minimum detonatable WD mass may be \(\simeq 0.85 \, M_\sun\).
## Spectra {#sec:spectra}
Figures [\[fig:spectra\]](#fig:spectra){reference-type="ref" reference="fig:spectra"} and [\[fig:spectra_3070\]](#fig:spectra_3070){reference-type="ref" reference="fig:spectra_3070"} compare synthetic near-maximum spectra of our five solar metallicity WD models with initial C/O mass fractions of \(50/50\) and \(30/70\), respectively, to spectra of SN 1999by, SN 2011fe, and SN1999dq at \(-2\), \(+0\), and \(\unit[+2]{d}\) from \(B\)-band maximum. While detailed features are not matched precisely, the overall agreement is promising. Our synthetic spectra show the hallmark attributes of SNe Ia--strong Si, S, Ca, and Fe features --with reasonable correspondence to observed line strengths. The \(\unit[4000-4500]{\AA}\) Ti [ii]{.smallcaps} trough characteristic of subluminous SNe Ia is also partially reproduced in our least massive \(0.8\) and \(0.85 \, M_\sun\) models. One of the most significant discrepancies between our synthetic spectra and observations are the IME velocities, particularly for the more massive WDs \(\ge 1.0 \, M_\sun\) and brighter observed SNe. At \(B\)-band maximum, the Si and Ca velocities of our \(1.0\) and \(1.1 \, M_\sun\) explosions are several thousand \(\unit}\) higher than observed. It is possible that non-LTE calculations will alleviate this discrepancy, as 's sub-\(M_{\rm Ch}\) spectra possess appropriate line velocities. A resolution to this issue may also lie in future multi-dimensional explosion studies of the double detonation scenario. In the converging shock variant of the double detonation, a helium shell detonation propagates around the WD's surface and launches an oblique shock into the core that focuses its energy near the center and ignites the carbon detonation. This inwardly propagating shock may tamp the outgoing core detonation somewhat and reduce the velocities of the outermost ejecta where the IME features form. In the edge-lit double detonation variation, the helium shell detonation transitions into a carbon-powered detonation as soon as it encounters the WD core. Thus, for one hemisphere of the WD, the carbon detonation actually moves inwards initially, so that when pressure forces cause the ejecta to rebound outwards, the outermost ejecta velocity will be similarly limited. For the opposite hemisphere, tamping of the outer ejecta may still occur if the helium shell detonation races ahead and reaches the opposite pole before the carbon detonation traverses the WD core. performed multi-dimensional converging shock double detonation simulations that are similar to our planned future calculations. They found that standard SN Ia light curves and spectra are only produced if the helium shells are heavily polluted by \(^{12}\)C (\(\sim 30\%\) by mass). However, the minimum detonatable helium shell masses found by are an order of magnitude smaller than those used by. We remain hopeful that these much smaller realistic helium shells will still lead to tamping of the bulk ejecta's velocities without adversely affecting the light curves and overall spectra. There is also the intriguing possibility that these minimal helium shells, which only produce Si and Ca ashes, will also explain the high-velocity (\(\gtrsim \unit[2\E{4}]{km \, s^{-1}}\)) features seen in most SNe Ia.
# Conclusions {#sec:conc}
Motivated by discrepancies in the literature and a need for detailed nucleosynthetic data, we have revisited simulations of bare sub-\(M_{\rm Ch}\) C/O WD detonations. We use a detonation-broadening scheme in a \(41\)-isotope hydrodynamical simulation to spatially resolve the detonation structure and show convergence of the results with increasing resolution. These results are then post-processed with a \(205\)-isotope nuclear reaction network. Our bulk nucleosynthetic results confirm recent work by and disagree with the studies by and, especially for low-mass WDs. Our examination of neutron-rich nucleosynthesis counters some of the previous claims for \(M_{\rm Ch}\) explosions from the solar abundance of Mn and from observations of SN remnants, but future work is necessary to resolve remaining tensions. The synthetic light curves and spectra of our simulations show promising similarities to observations. We find that typical SN 2011fe-like SNe Ia can be produced by the detonations of \(1.0 \, M_\sun\) WDs, which are more numerous than the \(1.1 \, M_\sun\) WDs required by to produce typical SNe Ia. This lower mass requirement will increase binary population synthesis rates of SNe Ia from the dynamically driven double degenerate double detonation scenario as well as from violent double WD mergers that directly ignite carbon. This revision of the mapping of detonating WD mass to SN Ia luminosity is a necessary input to recent work on the evolution of the SN Ia luminosity function, another piece of evidence that the bulk of SNe Ia arise from sub-\(M_{\rm Ch}\) WD explosions. The peak luminosities and evolutionary timescales of our radiative transfer results are correlated in a similar way to the observed relation, and the spectral features and line ratios are in general agreement with observed spectra. However, there is significant disagreement in the line centers of the IMEs and in the evolutionary timescales for the high-mass WD explosions. We are hopeful that future calculations building on this work, including more precise treatments of radiation transport, will resolve these discrepancies. We will also verify the yields produced by this front-broadening scheme using comparisons to fully-resolved calculations of the microscopic structure of steady-state detonations. Future work will also include multi-dimensional simulations with the very low-mass detonatable helium shells found by. We will employ a similar detonation-broadening scheme, which provides an artificial but numerically resolved model that we expect to give resolution-independent results for modest grid resolutions. The inward shock from the helium detonation has the potential to tamp the IME velocities and bring our radiative transfer results into agreement with observations, and the ashes from the helium-burning may also provide a satisfying explanation for the high-velocity features observed in most SNe Ia. While much future work remains to be done, this study has bolstered the potential for sub-\(M_{\rm Ch}\) WD detonations in double WD binaries to explain most SNe Ia. Theoretical and observational studies are beginning to converge, and we are hopeful that the solution to the SN Ia progenitor mystery now lies within reach. We thank Tony Piro for his encouragement, Stefan Taubenberger for providing access to data, and Carles Badenes, Stéphane Blondin, Hector Martínez-Rodríguez, Alison Miller, Rüdiger Pakmor, Stuart Sim, and Frank Timmes for helpful discussions. KJS, DMT, and BJM received support from the NASA Astrophysics Theory Program (NNX15AB16G and NNX17AG28G). DK is supported in part by a DOE Office of Nuclear Physics Early Career Award, and by the Director, Office of Energy Research, Office of High Energy and Nuclear Physics, Divisions of Nuclear Physics, of the US DOE under Contract No. DE-AC02-05CH11231. This research used the Savio computational cluster resource provided by the Berkeley Research Computing program at the University of California, Berkeley (supported by the UC Berkeley Chancellor, Vice Chancellor of Research, and Office of the CIO), high-performance computing resources provided by the University of Alabama, and resources of the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported by the Office of Science of the US DOE under Contract No. DE-AC02-05CH11231.
[^1]: `http://mesa.sourceforge.net`, version 8845; default options used unless otherwise noted.
[^2]: `http://flash.uchicago.edu`, version 4.2.2; default options used unless otherwise noted.
[^3]: Much of the data used in this work was obtained through `https://sne.space`. | {'timestamp': '2018-02-01T02:01:14', 'yymm': '1706', 'arxiv_id': '1706.01898', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01898'} |
# Abstract (Not appropriate in this style!) {#abstract-not-appropriate-in-this-style .unnumbered} | {'timestamp': '2017-07-20T02:04:39', 'yymm': '1706', 'arxiv_id': '1706.01584', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01584'} |
null | null |
null | null |
null | null |
null | null |
# Introduction {#sec:1}
Let \(C\) be a (binary) singly even self-dual code. All codes in this note are binary. Let \(C_0\) denote the subcode of \(C\) consisting of codewords having weight \(\equiv0\pmod4\). The *shadow* \(S\) of \(C\) is defined to be \(C_0^\perp \setminus C\). Shadows for self-dual codes were introduced by Conway and Sloane in order to derive new upper bounds for the minimum weight of singly even self-dual codes, and to provide restrictions on the weight enumerators of singly even self-dual codes. By considering shadows, the largest possible minimum weights of singly even self-dual codes of lengths up to \(72\) are given in . In this note, we say that a singly even self-dual code with the largest possible minimum weight given in is *extremal*. The possible weight enumerators of extremal singly even self-dual codes are given in for lengths up to \(64\) and length \(72\) (see also for length \(60\)). It is a fundamental problem to find which weight enumerators actually occur for the possible weight enumerators (see ). The possible weight enumerators of extremal singly even self-dual \([60,30,12]\) codes are known as follows: \[\begin{aligned}
W_{60,1} &=& 1 +(2555+64 \beta) y^{12} + (33600-384 \beta)y^{14} + \cdots, \\ W_{60,2} &=& 1 + 3451y^{12} + 24128 y^{14} + \cdots, \end{aligned}\] where \(\beta\) is an integer. If there is an extremal singly even self-dual \([60,30,12]\) code with weight enumerator \(W_{60,1}\), then \(\beta \in \{0,1,2,\ldots,8,10\}\) . For \(\beta=0,1,5,7\) and \(10\), an extremal singly even self-dual code with weight enumerator \(W_{60,1}\) was found in ,, , and, respectively. An extremal singly even self-dual code with weight enumerator \(W_{60,2}\) was found in . One of the main aims of this note is to show the following:
These codes are constructed from four-circulant singly even self-dual \([60,30,d]\) codes for \(d=10\) and \(12\) by considering self-dual neighbors. It remains to determine whether there is an extremal singly even self-dual \([60,30,12]\) code with weight enumerator \(W_{60,1}\) for \(\beta=3,4,8\). The possible weight enumerators of extremal singly even self-dual \([58,29,10]\) codes are known as follows: \[\begin{aligned}
W_{58,1} &=& 1 + (165-2 \gamma)y^{10} + (5078 + 2 \gamma)y^{12} + \cdots, \\ W_{58,2} &=& 1 +(319-24 \beta-2 \gamma)y^{10} + (3132 + 152 \beta + 2\gamma)y^{12} + \cdots, \end{aligned}\] where \(\beta,\gamma\) are integers . If there is an extremal singly even self-dual \([58,29,10]\) code with weight enumerator \(W_{58,2}\), then \(\beta \in \{0,1,2\}\) . An extremal singly even self-dual code with weight enumerator \(W_{58,1}\) is known for \(\gamma=55\) . An extremal singly even self-dual code with weight enumerator \(W_{58,2}\) is known for \[\begin{aligned}
\beta=0 \text{ and } & \gamma\in \{2m \mid m =0,1,5,6,8,9,10,11,13,\ldots,65,68,71,79\}, \\ \beta=1 \text{ and } & \gamma\in \{2m \mid m =13,14,16,\ldots,58,63\}, \\ \beta=2 \text{ and } & \gamma\in \{2m \mid m =0,16,\ldots,50,55\} \end{aligned}\] (see, , ). The following proposition is one of the main results of this note.
These codes are constructed from extremal singly even self-dual \([60,30,12]\) codes constructed in this note by subtracting and their self-dual neighbors. Finally, we give some restriction on the possible weight enumerators of certain singly even self-dual codes with shadow of minimum weight \(1\) (Proposition [\[prop:restriction\]](#prop:restriction){reference-type="ref" reference="prop:restriction"}). As a consequence, it is shown that \(\gamma=55\) for the possible weight enumerator \(W_{58,1}\) (Corollary [\[cor\]](#cor){reference-type="ref" reference="cor"}). All self-dual codes in this note are singly even. From now on, we omit the term singly even. All computer calculations in this note were done with the help of Magma .
# Extremal four-circulant self-dual \([60,30,12]\) codes {#sec:4cir}
An \(n \times n\) circulant matrix has the following form: \[\left( \begin{array}{ccccc} r_0&r_1&r_2& \cdots &r_{n-1} \\ r_{n-1}&r_0&r_1& \cdots &r_{n-2} \\ \vdots &\vdots & \vdots && \vdots\\ r_1&r_2&r_3& \cdots&r_0 \end{array} \right),\] so that each successive row is a cyclic shift of the previous one. Let \(A\) and \(B\) be \(n \times n\) circulant matrices. Let \(C\) be a \([4n,2n]\) code with generator matrix of the following form: \[\label{eq:GM} \left( \begin{array}{ccc@{}c} \quad & {\Large I_{2n}} & \quad & \begin{array}{cc} A & B \\ B^T & A^T \end{array} \end{array} \right),\] where \(I_n\) denotes the identity matrix of order \(n\) and \(A^T\) denotes the transpose of \(A\). It is easy to see that \(C\) is self-dual if \(AA^T+BB^T=I_n\). The codes with generator matrices of the form [\[eq:GM\]](#eq:GM){reference-type="eqref" reference="eq:GM"} are called *four-circulant*. In this section, we give a classification of extremal four-circulant self-dual \([60,30,12]\) codes. Two codes are *equivalent* if one can be obtained from the other by a permutation of coordinates. Our exhaustive search found all distinct extremal four-circulant self-dual \([60,30,12]\) codes, which must be checked further for equivalence to complete the classification. This was done by considering all pairs of \(15 \times 15\) circulant matrices \(A\) and \(B\) satisfying the condition that \(AA^T+BB^T=I_{15}\), the sum of the weights of the first rows of \(A\) and \(B\) is congruent to \(1 \pmod 4\) and the sum of the weights is greater than or equal to \(13\). Since a cyclic shift of the first rows gives an equivalent code, we may assume without loss of generality that the last entry of the first row of \(B\) is \(1\). Then our computer search shows that the above distinct extremal four-circulant self-dual \([60,30,12]\) codes are divided into \(13\) inequivalent codes.
We denote the \(13\) codes by \(C_{60,i}\) \((i=1,2,\ldots,13)\). For the \(13\) codes \(C_{60,i}\) \((i=1,2,\ldots,13)\), the first rows \(r_A\) (resp. \(r_B\)) of the circulant matrices \(A\) (resp. \(B\)) in generator matrices [\[eq:GM\]](#eq:GM){reference-type="eqref" reference="eq:GM"} are listed in Table [1](#Tab:60){reference-type="ref" reference="Tab:60"}. We verified that the codes \(C_{60,i}\) have weight enumerator \(W_{60,1}\), where \(\beta\) are also listed in Table [1](#Tab:60){reference-type="ref" reference="Tab:60"}.
We denote the \(113\) codes by \(G_{60,i}\) \((i=1,2,\ldots,113)\). For the \(13\) codes \(G_{60,i}\) \((i=1,2,\ldots,13)\), the first rows \(r_A\) (resp. \(r_B\)) of the circulant matrices \(A\) (resp. \(B\)) in generator matrices [\[eq:GM\]](#eq:GM){reference-type="eqref" reference="eq:GM"} are listed in Table [4](#Tab:d10){reference-type="ref" reference="Tab:d10"}. The first rows for the all codes can be obtained from <http://www.math.is.tohoku.ac.jp/~mharada/Paper/60-4cir-d10.txt>.
For example, Proposition [\[prop:restriction\]](#prop:restriction){reference-type="ref" reference="prop:restriction"} can be applied to the following parameters: \[(n,d)=(58,10), (74,14) \text{ and } (98,18).\]
- \((n,d)=(58,10)\): The possible weight enumerator of the shadow of an extremal self-dual \([58,29,10]\) code with weight enumerator \(W_{58,1}\) is as follows : \[y+ \gamma y^9 + (23918-10\gamma)y^{13} + \cdots.\] By Proposition [\[prop:restriction\]](#prop:restriction){reference-type="ref" reference="prop:restriction"}, we have \[165-2\gamma = \gamma.\] Since there is an extremal self-dual \([58,29,10]\) code with weight enumerator \(W_{58,1}\) for \(\gamma=55\) , we have the following:
::: {.cor}
[\[cor\]]{#cor label="cor"} There is an extremal self-dual \([58,29,10]\) code with weight enumerator \(W_{58,1}\) if and only if \(\gamma=55\).
:::
- \((n,d)=(74,14)\): The weight enumerator \(W_2\) in is the possible weight enumerator of an extremal self-dual \([74,37,14]\) code with shadow of minimum weight \(1\). By Proposition [\[prop:restriction\]](#prop:restriction){reference-type="ref" reference="prop:restriction"}, we have \(\alpha=-135\) for \(W_2\) in. The weight enumerators of such a code and its shadow are as follows: \[\begin{aligned}
&1 + 2044 y^{14} + 159067 y^{16} + 520782 y^{18} + \cdots, \\ &y + 2044 y^{13} + 679849 y^{17} + 44010824 y^{21} +\cdots, \end{aligned}\] respectively. It is still unknown whether there is an extremal self-dual \([74,37,14]\) code (with shadow of minimum weight \(1\)).
- \((n,d)=(98,18)\): The weight enumerator \(W_3\) in is the unique weight enumerator for an extremal self-dual \([98,49,18]\) code with shadow of minimum weight \(1\). The weight enumerators of such a code and its shadow are as follows: \[\begin{aligned}
&1 + 22116 y^{18} + 2016048 y^{20} + 7181104 y^{22} + \cdots, \\ &y + 22116 y^{17} + 9197152 y^{21} + 964758896 y^{25} + \cdots, \end{aligned}\] respectively. It is still unknown whether there is an extremal self-dual \([98,49,18]\) code (with shadow of minimum weight \(1\)).
This work was supported by JSPS KAKENHI Grant Number 15H03633. The author would like to thank the anonymous reviewers for the useful comments. | {'timestamp': '2017-06-07T02:05:21', 'yymm': '1706', 'arxiv_id': '1706.01694', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01694'} |
# Introduction
It is well known that solutions \(u\) of the \(p\)-Laplace equation, for \(1<p<\infty\), can be characterized as local minimizers of the \(L^p\)-norm of \(|\nabla u|\). This formulation has the advantage that it can be generalized to a metric measure space, by replacing \(|\nabla u|\) with the minimal \(p\)-weak upper gradient \(g_u\); see Section [2](#preliminaries){reference-type="ref" reference="preliminaries"} for definitions and notation. The study of such \(p\)-minimizers is a starting point for nonlinear potential theory, which is now well developed even in metric spaces that are equipped with a doubling measure and support a Poincaré inequality, see especially the monograph and e.g., and also the monographs and for the Euclidean theory and its history in the nonweighted and weighted setting, respectively. In the case \(p=1\), instead of the \(p\)-energy it is natural to minimize the total variation among functions of bounded variation (\(\mathrm{BV}\) functions), and the resulting minimizers are called functions of least gradient, see e.g. for previous works in the Euclidean setting, and in the metric setting. More precisely, a function \(u\in\mathrm{BV}_{\mathrm{loc}}(\Omega)\) is a function of least gradient in an open set \(\Omega\subset X\) if for every \(\varphi\in \mathrm{BV}_c(\Omega)\), we have \[\Vert Du\Vert(\supp \varphi)\le \Vert D(u+\varphi)\Vert(\supp \varphi)\] Testing only with nonnegative \(\varphi\) leads to the notion of *1-superminimizers*, whose study is the main objective of this paper. In the case \(p>1\), much of potential theory deals with superminimizers and the closely related concept of superharmonic functions, which were introduced in the metric setting in. A notion of \(1\)-superharmonic functions has been studied in the Euclidean setting in, but especially in the metric setting very little is known about these concepts in the case \(p=1\). For consistency, we use the term *1-minimizer* instead of function of least gradient. In, \(1\)-minimizers were shown to be continuous outside their jump sets. In this paper we show that this is a consequence of the fact that for super-and subminimizers \(u\), the pointwise representatives \(u^{\wedge}\) and \(u^{\vee}\) are lower and upper semicontinuous, respectively, at *every* point. This is Theorem [\[thm:superminimizers are lsc\]](#thm:superminimizers are lsc){reference-type="ref" reference="thm:superminimizers are lsc"}. We also study some basic properties of solutions of *obstacle problems* in the case \(p=1\); such solutions are, in particular, \(1\)-superminimizers. In the Euclidean setting, obstacle problems for the \(\mathrm{BV}\) class have been studied in e.g., and in the metric setting in. In this paper we first prove standard De Giorgi-type and weak Harnack inequalities for \(1\)-subminimizers and certain solutions of obstacle problems, following especially, and then use these to show the aforementioned semicontinuity property of \(1\)-superminimizers as well as a similar property for solutions of obstacle problems at points where the obstacle is continuous, see Theorem [\[thm:semicontinuity of obstacle problems\]](#thm:semicontinuity of obstacle problems){reference-type="ref" reference="thm:semicontinuity of obstacle problems"}. While these results are of some independent interest, our main goal is to consider certain questions of *fine potential theory* when \(p=1\). In the case \(p>1\) it is known that the so-called *\(p\)-fine topology* is the coarsest topology that makes all \(p\)-superharmonic functions continuous in open sets (alternatively upper semicontinuous, as superharmonic functions are lower semicontinuous already with respect to the metric topology). In Section [4](#sec:the fine topology){reference-type="ref" reference="sec:the fine topology"} we define the notion of *thinness* and the resulting fine topology in the case \(p=1\), following, and generalize some properties concerning, in particular, points of nonzero capacity from the case \(p>1\) to the case \(p=1\). Then in Section [5](#sec:weak cartan property){reference-type="ref" reference="sec:weak cartan property"} we prove the main result of this paper, namely the following *weak Cartan property* for \(1\)-superminimizers.
The analogous property in the case \(p>1\) is well known and proved in the metric setting in. We also prove a strong version of the property, requiring only one superminimizer, at points of nonzero capacity; this is Proposition [\[prop:strong Cartan property\]](#prop:strong Cartan property){reference-type="ref" reference="prop:strong Cartan property"}. Then as in the case \(p>1\) we use the weak Cartan property to show that any topology that makes the upper approximate limit \(u^{\vee}\) of every \(1\)-superminimizer \(u\) upper semicontinuous in every open set \(\Omega\) necessarily contains the \(1\)-fine topology. This is Theorem [\[thm:superminimizer top contains fine top\]](#thm:superminimizer top contains fine top){reference-type="ref" reference="thm:superminimizer top contains fine top"}. However, we observe that unlike in the case \(p>1\), the converse does not hold, that is, the \(1\)-fine topology does not always make \(u^{\vee}\) upper semicontinuous for \(1\)-superminimizers \(u\); see Example [\[ex:counterexample\]](#ex:counterexample){reference-type="ref" reference="ex:counterexample"}. Our main results seem to be new even in Euclidean spaces. A key motivation for the work is that a weak Cartan property will be useful in considering further questions such as *p-strict subsets*, fine connectedness, and the relationship between finely open and *quasiopen* sets for \(p=1\), see for example for the case \(p>1\).
# Preliminaries
In this section we introduce most of the notation, definitions, and assumptions employed in the paper. Throughout this paper, \((X,d,\mu)\) is a complete metric space that is equipped with a metric \(d\) and a Borel regular outer measure \(\mu\) that satisfies a doubling property. The doubling property means that there is a constant \(C_d\ge 1\) such that \[0<\mu(B(x,2r))\le C_d\mu(B(x,r))<\infty\] for every ball \(B(x,r):=\{y\in X:\,d(y,x)<r\}\) with center \(x\in X\) and radius \(r>0\). Sometimes we abbreviate \(B:=B(x,r)\) and \(aB:=B(x,ar)\) with \(a>0\); note that in metric spaces, a ball does not necessarily have a unique center point and radius, but we will always consider balls for which these have been specified. We also assume that \(X\) supports a \((1,1)\)-Poincaré inequality that will be defined below, and that \(X\) consists of at least \(2\) points. By iterating the doubling condition, we obtain for any \(x\in X\) and any \(y\in B(x,R)\) with \(0<r\le R<\infty\) that \[\label{eq:homogenous dimension} \frac{\mu(B(y,r))}{\mu(B(x,R))}\ge \frac{1}{C_d^2}\left(\frac{r}{R}\right)^{Q},\] where \(Q>1\) only depends on the doubling constant \(C_d\). When we want to state that a constant \(C\) depends on the parameters \(a,b, \ldots\), we write \(C=C(a,b,\ldots)\), and we understand all constants to be strictly positive. When a property holds outside a set of \(\mu\)-measure zero, we say that it holds almost everywhere, abbreviated a.e. A complete metric space equipped with a doubling measure is proper, that is, closed and bounded sets are compact. Since \(X\) is proper, for any open set \(\Omega\subset X\) we define \(\mathrm{Lip}_{\mathrm{loc}}(\Omega)\) to be the space of functions that are Lipschitz in every open \(\Omega'\Subset\Omega\). Here \(\Omega'\Subset\Omega\) means that \(\overline{\Omega'}\) is a compact subset of \(\Omega\). Other local spaces of functions are defined analogously. For any set \(A\subset X\) and \(0<R<\infty\), the restricted spherical Hausdorff content of codimension one is defined to be \[\mathcal{H}_{R}(A):=\inf\left\{ \sum_{i=1}^{\infty} \frac{\mu(B(x_{i},r_{i}))}{r_{i}}:\,A\subset\bigcup_{i=1}^{\infty}B(x_{i},r_{i}),\,r_{i}\le R\right\}.\] The codimension one Hausdorff measure of \(A\subset X\) is then defined to be \[\mathcal{H}(A):=\lim_{R\rightarrow 0}\mathcal{H}_{R}(A).\] The measure theoretic boundary \(\partial^{*}E\) of a set \(E\subset X\) is the set of points \(x\in X\) at which both \(E\) and its complement have strictly positive upper density, i.e. \[\limsup_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}>0\quad \textrm{and}\quad\limsup_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}>0.\] The measure theoretic interior and exterior of \(E\) are defined respectively by \[\label{eq:definition of measure theoretic interior} I_E:=\left\{x\in X:\,\lim_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}=0\right\}\] and \[\label{eq:definition of measure theoretic exterior} O_E:=\left\{x\in X:\,\lim_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}=0\right\}.\] Note that the space is always partitioned into the disjoint sets \(\partial^*E\), \(I_E\), and \(O_E\). By Lebesgue's differentiation theorem (see e.g. ), for a \(\mu\)-measurable set \(E\) we have \(\mu(E\Delta I_E)=0\), where \(\Delta\) is the symmetric difference. All functions defined on \(X\) or its subsets will take values in \(\overline{{\mathbb R}}:=[-\infty,\infty]\). By a curve we mean a rectifiable continuous mapping from a compact interval of the real line into \(X\). A nonnegative Borel function \(g\) on \(X\) is an upper gradient of a function \(u\) on \(X\) if for all nonconstant curves \(\gamma\), we have \[\label{eq:definition of upper gradient} |u(x)-u(y)|\le \int_\gamma g\,ds,\] where \(x\) and \(y\) are the end points of \(\gamma\) and the curve integral is defined by using an arc-length parametrization, see where upper gradients were originally introduced. We interpret \(|u(x)-u(y)|=\infty\) whenever at least one of \(|u(x)|\), \(|u(y)|\) is infinite. In what follows, let \(1\le p<\infty\). We say that a family of curves \(\Gamma\) is of zero \(p\)-modulus if there is a nonnegative Borel function \(\rho\in L^p(X)\) such that for all curves \(\gamma\in\Gamma\), the curve integral \(\int_\gamma \rho\,ds\) is infinite. A property is said to hold for \(p\)-almost every curve if it fails only for a curve family with zero \(p\)-modulus. If \(g\) is a nonnegative \(\mu\)-measurable function on \(X\) and ([\[eq:definition of upper gradient\]](#eq:definition of upper gradient){reference-type="ref" reference="eq:definition of upper gradient"}) holds for \(p\)-almost every curve, we say that \(g\) is a \(p\)-weak upper gradient of \(u\). By only considering curves \(\gamma\) in \(\Omega\subset X\), we can talk about a function \(g\) being a (\(p\)-weak) upper gradient of \(u\) in \(\Omega\). Given an open set \(\Omega\subset X\), we define the norm \[\Vert u\Vert_{N^{1,p}(\Omega)}:=\Vert u\Vert_{L^p(\Omega)}+\inf \Vert g\Vert_{L^p(\Omega)},\] where the infimum is taken over all \(p\)-weak upper gradients \(g\) of \(u\) in \(\Omega\). The substitute for the Sobolev space \(W^{1,p}\) in the metric setting is the Newton-Sobolev space \[N^{1,p}(\Omega):=\{u:\|u\|_{N^{1,p}(\Omega)}<\infty\}.\] We understand every Newton-Sobolev function to be defined at every \(x\in \Omega\) (even though \(\Vert \cdot\Vert_{N^{1,p}(\Omega)}\) is, precisely speaking, then only a seminorm). It is known that for any \(u\in N_{\mathrm{loc}}^{1,p}(\Omega)\), there exists a minimal \(p\)-weak upper gradient of \(u\) in \(\Omega\), always denoted by \(g_{u}\), satisfying \(g_{u}\le g\) a.e. in \(\Omega\), for any \(p\)-weak upper gradient \(g\in L_{\mathrm{loc}}^{p}(\Omega)\) of \(u\) in \(\Omega\), see. The \(p\)-capacity of a set \(A\subset X\) is given by \[\capa_p(A):=\inf \Vert u\Vert_{N^{1,p}(X)},\] where the infimum is taken over all functions \(u\in N^{1,p}(X)\) such that \(u\ge 1\) in \(A\). We know that \(\capa_p\) is an outer capacity, meaning that \[\capa_p(A)=\inf\{\capa_p(U):\,U\supset A\textrm{ is open}\}\] for any \(A\subset X\), see e.g.. If a property holds outside a set \(A\subset X\) with \(\capa_p(A)=0\), we say that it holds \(p\)-quasieverywhere, abbreviated \(p\)-q.e. If \(u\in N^{1,p}(\Omega)\), then \(\Vert u-v\Vert_{N^{1,p}(\Omega)}=0\) if and only if \(u=v\) \(p\)-q.e. in \(\Omega\), see. By we know that if \(A\subset X\), \[\label{eq:null sets of Hausdorff measure and capacity} \capa_1(A)=0\quad\textrm{if and only if}\quad\mathcal H(A)=0.\] The variational \(p\)-capacity of a set \(A\subset \Omega\) with respect to an open set \(\Omega\subset X\) is given by \[\rcapa_p(A,\Omega):=\inf \int_X g_u^p \,d\mu,\] where the infimum is taken over functions \(u\in N^{1,p}(X)\) such that \(u\ge 1\) in \(A\) (equivalently, \(p\)-q.e. in \(A\)) and \(u=0\) in \(X\setminus \Omega\); recall that \(g_u\) is the minimal \(p\)-weak upper gradient of \(u\). We know that \(\rcapa_p\) is also an outer capacity, in the sense that if \(\Omega\subset X\) is a bounded open set and \(A\Subset \Omega\), then \[\rcapa_p(A,\Omega)=\inf\{\rcapa_p(U,\Omega):\,U\textrm{ open},\,A\subset U\subset\Omega \},\] see. It is easy to see that in the definitions of capacities, we can assume the test functions to satisfy \(0\le u\le 1\). For basic properties satisfied by capacities, such as monotonicity and countable subadditivity, see e.g.. Next we recall the definition and basic properties of functions of bounded variation on metric spaces, following. See also e.g. for the classical theory in the Euclidean setting. Let \(\Omega\subset X\) be an open set. Given a function \(u\in L^1_{\mathrm{loc}}(\Omega)\), we define the total variation of \(u\) in \(\Omega\) by \[\|Du\|(\Omega):=\inf\left\{\liminf_{i\to\infty}\int_{\Omega} g_{u_i}\,d\mu:\, u_i\in \Lip_{\mathrm{loc}}(\Omega),\, u_i\to u\textrm{ in } L^1_{\mathrm{loc}}(\Omega)\right\},\] where each \(g_{u_i}\) is again the minimal \(1\)-weak upper gradient of \(u_i\) in \(\Omega\). (In, local Lipschitz constants were used instead of upper gradients, but the properties of the total variation can be proved similarly with either definition.) We say that a function \(u\in L^1(\Omega)\) is of bounded variation, and denote \(u\in\mathrm{BV}(\Omega)\), if \(\|Du\|(\Omega)<\infty\). For an arbitrary set \(A\subset X\), we define \[\|Du\|(A):=\inf\{\|Du\|(U):\, A\subset U,\,U\subset X \text{ is open}\}.\] If \(u\in L^1_{\mathrm{loc}}(\Omega)\) and \(\Vert Du\Vert(\Omega)<\infty\), \(\|Du\|(\cdot)\) is a Radon measure on \(\Omega\) by. A \(\mu\)-measurable set \(E\subset X\) is said to be of finite perimeter if \(\|D\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_E\|(X)<\infty\), where \(\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_E\) is the characteristic function of \(E\). The perimeter of \(E\) in \(\Omega\) is also denoted by \[P(E,\Omega):=\|D\text{\raise 1.3pt \hbox{\(\chi\)}\kern-0.2pt}_E\|(\Omega).\] For any \(u,v\in L^1_{\mathrm{loc}}(\Omega)\), it is straightforward to show that \[\label{eq:BV functions form vector space} \Vert D(u+v)\Vert(\Omega)\le \Vert Du\Vert(\Omega)+\Vert Dv\Vert(\Omega).\] We have the following coarea formula from: if \(\Omega\subset X\) is an open set and \(u\in L^1_{\mathrm{loc}}(\Omega)\), then \[\label{eq:coarea} \|Du\|(\Omega)=\int_{-\infty}^{\infty}P(\{u>t\},\Omega)\,dt.\] We will assume throughout the paper that \(X\) supports a \((1,1)\)-Poincaré inequality, meaning that there exist constants \(C_P>0\) and \(\lambda \ge 1\) such that for every ball \(B(x,r)\), every \(u\in L^1_{\mathrm{loc}}(X)\), and every upper gradient \(g\) of \(u\), we have \[\vint{B(x,r)}|u-u_{B(x,r)}|\, d\mu \le C_P r\vint{B(x,\lambda r)}g\,d\mu,\] where \[u_{B(x,r)}:=\vint{B(x,r)}u\,d\mu :=\frac 1{\mu(B(x,r))}\int_{B(x,r)}u\,d\mu.\] Applying the Poincaré inequality to sequences of approximating locally Lipschitz functions in the definition of the total variation gives the following \(\mathrm{BV}\) version: for every ball \(B(x,r)\) and every \(u\in L^1_{\mathrm{loc}}(X)\), we have \[\vint{B(x,r)}|u-u_{B(x,r)}|\,d\mu \le C_P r\, \frac{\Vert Du\Vert (B(x,\lambda r))}{\mu(B(x,\lambda r))}.\] For a \(\mu\)-measurable set \(E\subset X\), the above implies (see e.g. ) the relative isoperimetric inequality \[\label{eq:relative isoperimetric inequality} \min\{\mu(B(x,r)\cap E),\,\mu(B(x,r)\setminus E)\}\le 2 C_P rP(E,B(x,\lambda r)).\] Moreover, from the \((1,1)\)-Poincaré inequality, by we get the following Sobolev inequality: if \(x\in X\), \(0<r<\frac{1}{4}\diam X\), and \(u\in N^{1,1}(X)\) with \(u=0\) in \(X\setminus B(x,r)\), then \[\label{eq:sobolev inequality} \left(\,\vint{B(x,r)} |u|^{Q/(Q-1)}\,d\mu\right)^{(Q-1)/Q} \le C_S r \vint{B(x,r)} g_u\,d\mu\] for a constant \(C_S=C_S(C_d,C_P,\lambda)\ge 1\). Then for any \(x\in X\), any \(0<r<\frac{1}{4}\diam X\), and any \(u\in L^1_{\mathrm{loc}}(X)\) with \(u=0\) in \(X\setminus B(x,r)\), by applying the above to a suitable sequence approximating \(u\), we obtain \[\label{eq:sobolev inequality for BV} \left(\,\vint{B(x,r)} |u|^{Q/(Q-1)}\,d\mu\right)^{(Q-1)/Q} \le C_S r \frac{\Vert Du\Vert(X)}{\mu(B(x,r))}.\] For any \(\mu\)-measurable set \(E\subset B(x,r)\), this implies by Hölder's inequality \[\label{eq:isop inequality with zero boundary values} \mu(E)\le C_S r P(E,X).\] Moreover, if \(\Omega\subset X\) is an open set with \(\diam \Omega<\frac 14 \diam X\) (meaning \(\diam \Omega<\infty\) in the case \(\diam X=\infty\)) and \(u\in L^1_{\mathrm{loc}}(X)\) with \(u=0\) in \(X\setminus\Omega\), then we can take a ball \(B(x,r)\supset \Omega\) with \(r= \diam \Omega\), and so by [\[eq:sobolev inequality for BV\]](#eq:sobolev inequality for BV){reference-type="eqref" reference="eq:sobolev inequality for BV"} and Hölder's inequality \[\label{eq:Poincare ineq for BV with zero boundary values} \int_{\Omega} |u|\,d\mu\le C_{S}\diam \Omega \Vert Du\Vert(X).\] The lower and upper approximate limits of a function \(u\) on \(X\) are defined respectively by \[\label{eq:lower approximate limit} u^{\wedge}(x): =\sup\left\{t\in{\mathbb R}:\,\lim_{r\to 0}\frac{\mu(\{u<t\}\cap B(x,r))}{\mu(B(x,r))}=0\right\}\] and \[\label{eq:upper approximate limit} u^{\vee}(x): =\inf\left\{t\in{\mathbb R}:\,\lim_{r\to 0}\frac{\mu(\{u>t\}\cap B(x,r))}{\mu(B(x,r))}=0\right\}.\] Unlike Newton-Sobolev functions, we understand \(\mathrm{BV}\) functions to be \(\mu\)-equivalence classes. To consider fine properties, we need to consider the pointwise representatives \(u^{\wedge}\) and \(u^{\vee}\).
# Superminimizers and obstacle problems
In this section we consider superminimizers and solutions of obstacle problems in the case \(p=1\). The symbol \(\Omega\) will always denote a nonempty open subset of \(X\). We denote by \(\mathrm{BV}_c(\Omega)\) the class of functions \(\varphi\in\mathrm{BV}(\Omega)\) with compact support in \(\Omega\), that is, \(\supp \varphi\Subset \Omega\).
Equivalently, we can replace \(\supp\varphi\) by any set \(A\Subset \Omega\) containing \(\supp\varphi\) in the above definitions. It is easy to see that if \(u\) is a \(1\)-superminimizer and \(a\ge 0\), \(b\in {\mathbb R}\), then \(au+b\) is a \(1\)-superminimizer. Given a nonempty bounded open set \(\Omega\subset X\), a function \(\psi\colon\Omega\to\overline{{\mathbb R}}\), and \(f\in L^1_{\mathrm{loc}}(X)\) with \(\Vert Df\Vert(X)<\infty\), we define the class of admissible functions \[\mathcal K_{\psi,f}(\Omega):=\{u\in\mathrm{BV}_{\mathrm{loc}}(X):\,u\ge \psi\textrm{ in }\Omega\textrm{ and }u=f\textrm{ in }X\setminus\Omega\}.\] The (in)equalities above are understood in the a.e. sense, since \(\mathrm{BV}\) functions are only defined up to sets of \(\mu\)-measure zero. For brevity, we sometimes write \(\mathcal K_{\psi,f}\) instead of \(\mathcal K_{\psi,f}(\Omega)\). By using a cutoff function, it is easy to show that \(\Vert Du\Vert(X)<\infty\) for every \(u\in\mathcal K_{\psi,f}(\Omega)\).
Unlike in the case \(p>1\), solutions are not generally unique, as can be easily seen for example by considering translates of the Heaviside function on the real line. The following fact, which is also in stark contrast to the case \(p>1\), is often useful.
Whenever the characteristic function of a set \(E\) is a solution of an obstacle problem, for simplicity we will call \(E\) a solution as well. Similarly, if \(\psi=\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_A\) for some \(A\subset X\), we let \(\mathcal K_{A, f}:=\mathcal K_{\psi, f}\). The following simple fact will be of much use to us.
The following fact follows directly from the definitions.
Next we prove De Giorgi-type and weak Harnack inequalities for \(1\)-subminimizers and certain solutions of obstacle problems. The arguments we use are mostly standard and have been employed in the metric setting previously in, but only for (quasi)minimizers or in the case \(p>1\), so we repeat the entire proofs with small modifications and some simplifications.
In proving the following weak Harnack inequality, we closely follow, where the analogous result is proved in the case \(p>1\). Recall the definition of the exponent \(Q>1\) from [\[eq:homogenous dimension\]](#eq:homogenous dimension){reference-type="eqref" reference="eq:homogenous dimension"}.
We combine the previous two propositions to get the following theorem. Recall that \(\Omega\) always denotes a nonempty open set.
Unlike \(p\)-harmonic functions for \(p>1\), \(1\)-minimizers are not always continuous with any choice of representative, as demonstrated already by the Heaviside function on the real line. However, the following semicontinuity holds. Recall the definitions of the pointwise representatives \(u^{\wedge}\) and \(u^{\vee}\) from [\[eq:lower approximate limit\]](#eq:lower approximate limit){reference-type="eqref" reference="eq:lower approximate limit"} and [\[eq:upper approximate limit\]](#eq:upper approximate limit){reference-type="eqref" reference="eq:upper approximate limit"}.
We conclude that for a \(1\)-minimizer \(u\), \(u^{\wedge}\) is lower semicontinuous and \(u^{\vee}\) is upper semicontinuous. From this we immediately get the following corollary, which was previously proved in. We define the jump set \(S_u\) of a \(\mathrm{BV}\) function \(u\) as the set where \(u^{\wedge}<u^{\vee}\).
For obstacle problems in the case \(p>1\), it is well known that continuity of the obstacle implies continuity of the solution, see or. In the case \(p=1\), the best we can hope for is lower semicontinuity of \(u^{\wedge}\) (which holds for superminimizers and thus for solutions of obstacle problems) and the upper semicontinuity of \(u^{\vee}\). These we can indeed obtain.
Here \(\esslimsup_{y\to x}\psi(y):=\lim_{r\to 0}\esssup_{B(x,r)}\psi\). In particular, it is enough if \(\psi\) is continuous (as an \(\overline{{\mathbb R}}\)-valued function) at \(x\). In Example [\[ex:counterexample\]](#ex:counterexample){reference-type="ref" reference="ex:counterexample"} we will see that \(u^{\vee}\) is not always upper semicontinuous.
For *general* \(\mathrm{BV}\) functions we have the following result, which follows from, and was proved earlier in the Euclidean setting in.
This *quasi-semicontinuity* is to be compared with the quasicontinuity of Newton-Sobolev functions: if \(u\in N^{1,p}(X)\) for \(1\le p<\infty\) and \(\varepsilon>0\), then there exists an open set \(G\subset X\) such that \(\capa_p(G)<\varepsilon\) and \(u|_{X\setminus G}\) is continuous; see or.
# The \(1\)-fine topology {#sec:the fine topology}
In this section we consider some basic properties of the *\(1\)-fine topology*. The following definition is from.
See for a proof of the fact that the \(1\)-fine topology is indeed a topology. We record the following fact given in, and use it to prove two lemmas that will be useful later.
The following is a standard result in the case \(p>1\), see e.g. or, and we prove it similarly for \(p=1\).
The analog of the next proposition is again known for \(p>1\), see, but in this case our proof will be rather different. In the case \(p>1\) the proof relies on the theory of \(p\)-harmonic functions, but we are able to use a more direct argument that relies on the relative isoperimetric inequality.
Towards proving the proposition, we first collect some more facts. According to, if \(x\in X\), \(0<r<\frac 18\diam X\), and \(A\subset B(x,r)\), then for some constant \(C=C(C_d,C_P,\lambda)\), \[\label{eq:comparison of capacities} \frac{\capa_1(A)}{C(1+r)}\le\rcapa_1(A,B(x,2r))\le 2\left(1+\frac 1r\right)\capa_1(A).\] In fact, the proof reveals that the second inequality holds with any \(r>0\). We will need one more estimate for the variational \(1\)-capacity; recall the definition of the measure theoretic interior \(I_A\) from [\[eq:definition of measure theoretic interior\]](#eq:definition of measure theoretic interior){reference-type="eqref" reference="eq:definition of measure theoretic interior"}.
We also need the following simple lemma.
Now we derive a converse type of result compared with Lemma [\[lem:solutions from capacity\]](#lem:solutions from capacity){reference-type="ref" reference="lem:solutions from capacity"}, given in Lemma [\[lem:capacity estimated by perimeter\]](#lem:capacity estimated by perimeter){reference-type="ref" reference="lem:capacity estimated by perimeter"} below.
We can strengthen this in the following way.
Finally, we record the following consequence of.
In other words, for \(1\)-q.e. \(x\in X\), every set \(\{u^{\wedge}>t\}\) (with \(t\in{\mathbb R}\)) that contains \(x\) is a \(1\)-fine neighborhood of \(x\). For Newton-Sobolev functions we have the stronger result that if \(u\in N^{1,p}(X)\) for \(1< p<\infty\), then \(u\) is \(p\)-finely continuous at \(p\)-q.e. \(x\in X\), see, , or; we do not give the definition of the \(p\)-fine topology for \(p>1\) here but it can also be found in the above references.
# The weak Cartan property {#sec:weak cartan property}
In this section we prove the weak Cartan property, as well as a strong version at points of nonzero \(1\)-capacity. Our proof will rely on breaking the set \(A\) into two subsets that do not intersect certain annuli around \(x\). Such a separation argument is inspired by the proof of the analogous property in the case \(p>1\), see, which in turn is based on and.
Now we prove the weak Cartan property, Theorem [\[thm:weak Cartan property\]](#thm:weak Cartan property){reference-type="ref" reference="thm:weak Cartan property"}. In fact, we give the following formulation containing somewhat more information, which will be useful in future work when considering *p-strict subsets* and a *Choquet property* in the case \(p=1\), cf., , and.
It can be noted that in the case \(p>1\), the proof of the weak Cartan property relies on the comparison principle as well as weak Harnack inequalities for both superminimizers and subminimizers. We only have the last of these three tools available, but we are able to replace the others (and in fact get a simpler argument) with the very powerful fact that the superminimizer functions can be taken to be characteristic functions of sets of finite perimeter; recall especially [\[eq:disconnectedness of Ei\]](#eq:disconnectedness of Ei){reference-type="eqref" reference="eq:disconnectedness of Ei"}.
The analog of the following result is again known in the case \(p>1\), see. Our proof will be similar, but we need to rely on the quasisemicontinuity of \(\mathrm{BV}\) functions instead of the quasicontinuity that is available in the case \(p>1\).
Note that this does not follow directly from the definition of \(1\)-thinness, since it is possible that \(r/\mu(B(x,r))\to 0\) as \(r\to 0\), recall Example [\[ex:singular points\]](#ex:singular points){reference-type="ref" reference="ex:singular points"}.
Just as in the case \(p>1\), see, at points of nonzero capacity we obtain a strong Cartan property, where we need only one superminimizer.
In the case \(p>1\), the \(p\)-fine topology is known to be the coarsest topology that makes all \(p\)-superharmonic functions on open subsets of \(X\) continuous, see. Equivalently, it is the coarsest topology that makes such functions upper semicontinuous, since they are lower semicontinuous already with respect to the metric topology. In the following we consider what the analog of this could be in the case \(p=1\).
Note that if \(X\) is bounded and thus compact, the only \(1\)-superminimizers in \(X\) are constants (for nonconstant \(u\in\mathrm{BV}(X)\) we have \(\Vert D\max\{u,k\}\Vert(X)<\Vert Du\Vert(X)\) for some \(k\in{\mathbb R}\)). This is why we want to talk about \(1\)-superminimizers in open sets \(\Omega\), and as a result, the metric topology is contained in the \(1\)-superminimizer topology by definition.
Now it might seem reasonable to postulate that the converse would hold as well, i.e. that the \(1\)-fine topology would make \(u^{\vee}\) upper semicontinuous for all \(1\)-superminimizers \(u\) in open sets. However, this is not the case.
Nevertheless, it is perhaps interesting to note that in Theorem [\[thm:weak Cartan property in text\]](#thm:weak Cartan property in text){reference-type="ref" reference="thm:weak Cartan property in text"}, \(\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_{E_0}^{\vee}\) and \(\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_{E_1}^{\vee}\) *are* \(1\)-finely upper semicontinuous at \(x\), since \(\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_{E_0}^{\vee}(x)=0=\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_{E_1}^{\vee}(x)\) and the sets \(\{\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_{E_0}^{\vee}>0\}\) and \(\{\text{\raise 1.3pt \hbox{\)\chi\(}\kern-0.2pt}_{E_1}^{\vee}>0\}\) are \(1\)-thin at \(x\). We expect this fact to be a useful substitute for fine upper semicontinuity in future research. In Table [1](#tab:comparison){reference-type="ref" reference="tab:comparison"} we compare the properties of Newton-Sobolev and \(p\)-superharmonic functions (for \(1<p<\infty\)) with the analogous properties of \(\mathrm{BV}\) functions and \(1\)-superminimizers. For the results in the left column, see the comment after Proposition [\[prop:quasisemicontinuity of BV\]](#prop:quasisemicontinuity of BV){reference-type="ref" reference="prop:quasisemicontinuity of BV"}, the comment after Theorem [\[thm:fine semicontinuity of BV functions\]](#thm:fine semicontinuity of BV functions){reference-type="ref" reference="thm:fine semicontinuity of BV functions"}, or, and. For the results in the right column, see Proposition [\[prop:quasisemicontinuity of BV\]](#prop:quasisemicontinuity of BV){reference-type="ref" reference="prop:quasisemicontinuity of BV"}, Theorem [\[thm:fine semicontinuity of BV functions\]](#thm:fine semicontinuity of BV functions){reference-type="ref" reference="thm:fine semicontinuity of BV functions"}, Theorem [\[thm:superminimizers are lsc\]](#thm:superminimizers are lsc){reference-type="ref" reference="thm:superminimizers are lsc"}, and Theorem [\[thm:superminimizer top contains fine top\]](#thm:superminimizer top contains fine top){reference-type="ref" reference="thm:superminimizer top contains fine top"}. | {'timestamp': '2018-01-29T02:07:58', 'yymm': '1706', 'arxiv_id': '1706.01873', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01873'} |
# Introduction
With the explosive growth of informal electronic communications such as email, social media, web comments, etc., colloquial languages that were historically unwritten are starting to be written for the first time. For these languages, there are extremely limited (approximately zero) resources available, not even large amounts of monolingual text data or possibly not even small amounts of monolingual text data. Even when audio resources are available, difficulties arise when converting sound to text. Moreover, the text data that can be obtained often has non-standard spellings and substantial code-switching with other traditionally written languages. In this paper we present a method for the acquisition of translation lexicons via loanwords and expert knowledge that requires zero resources of the borrowing language. Many historically unwritten languages borrow from highly resourced languages. Also, it is often feasible to locate a language expert to find out how sounds in these languages would be rendered if they were to be written as many of them are beginning to be written in social media, etc. We thus expect the general method to be applicable for multiple historically unwritten languages. In this paper we investigate inducing a Moroccan Darija-English translation lexicon via borrowed French words. Moroccan Darija is an historically unwritten dialect of Arabic spoken by millions but lacking in standardization and linguistic resources. Moroccan Darija is known to borrow many words from French, one of the most highly resourced languages in the world. By mapping Moroccan Darija-French borrowings to their donor French words, we can rapidly create lexical resources for portions of Moroccan Darija vocabulary for which no resources currently exist. For example, we could use one of many bilingual French-English dictionaries to bridge into English and create a Moroccan Darija-English translation lexicon that can be used to assist professional translation of Moroccan Darija into English and to assist with construction of Moroccan Darija-English Machine Translation (MT) systems. The rest of this paper is structured as follows. Section [2](#related){reference-type="ref" reference="related"} summarizes related work; section [3](#method){reference-type="ref" reference="method"} explains our method; section [4](#experiments){reference-type="ref" reference="experiments"} discusses experimental results of applying our method to the case of building a Moroccan Darija-English translation lexicon; and section [5](#conclusions){reference-type="ref" reference="conclusions"} concludes.
# Related Work {#related}
Translation lexicons are a core resource used for multilingual processing of languages. Manual creation of translation lexicons by lexicographers is time-consuming and expensive. There are more than 7000 languages in the world, many of which are historically unwritten. For a relatively small number of these languages there are extensive resources available that have been manually created. It has been noted by others that languages are organized into families and that using cognates between sister languages can help rapidly create translation lexicons for lower-resourced languages. For example, the methods in are able to detect that English *kilograms* maps to Portuguese *quilogramas* via bridge Spanish *kilogramos*. This general idea has been worked on extensively in the context of cognates detection, with 'cognate' typically re-defined to include loanwords as well as true cognates. The methods use monolingual data at a minimum and many signals such as orthographic similarity, phonetic similarity, contextual similarity, temporal similarity, frequency similarity, burstiness similarity, and topic similarity . Inducing translations via loanwords was specifically targeted in. While some of these methods don't require bilingual resources, with the possible exception of small bilingual seed dictionaries, they do at a minimum require monolingual text data in the languages to be modeled and sometimes have specific requirements on the monolingual text data such as having text coming from the same time period for each of the languages being modeled. For colloquial languages that were historically unwritten, but that are now starting to be written with the advent of social media and web comments, there are often extremely limited resources of any type available, not even large amounts of monolingual text data. Moreover, the written data that can be obtained often has non-standard spellings and code-switching with other traditionally written languages. Often the code-switching occurs within words whereby the base is borrowed and the affixes are not borrowed, analogous to the multi-language categories "V\" and "N\" from. The data available for historically unwritten languages, and especially the lack thereof, is not suitable for previously developed cognates detection methods that operate as discussed above. In the next section we present a method for translation lexicon induction via loanwords that uses expert knowledge and requires zero resources from the borrowing language other than a language informant.
# Method
Our method is to take word pronunciations from the donor language we are using and convert them to how they would be rendered in the borrowing language if they were to be borrowed. These are our candidate loanwords. There are three possible cases for a given generated candidate loanword string:
true match
: string occurs in borrowing language and is a loanword from the donor language;
false match
: string occurs in borrowing language by coincidence but it's not a loanword from the donor language;
no match
: string does not occur in borrowing language.
For the case of inducing a Moroccan Darija-English translation lexicon via French we start with a French-English bilingual dictionary and take all the French pronunciations in IPA (International Phonetic Alphabet)[^1] and convert them to how they would be rendered in Arabic script. For this we created a multiple-step transliteration process:
Step 1
: Break pronunciation into syllables.
Step 2
: Convert each IPA syllable to a string in modified Buckwalter transliteration[^2], which supports a one-to-one mapping to Arabic script.
Step 3
: Convert each syllable's string in modified Buckwalter transliteration to Arabic script.
Step 4
: Merge the resulting Arabic script strings for each syllable to generate a candidate loanword string.
For syllabification, for many word pronunciations the syllables are already marked in the IPA by the '.' character; if syllables are not already marked in the IPA, we run a simple syllabifier to complete step 1. For step 2, we asked a language expert to give us a sequence of rules to convert a syllable's pronunciation to modified Buckwalter transliteration. This is itself a multi-step process (see next paragraph for details). In step 3, we simply do the one-to-one conversion and obtain Arabic script for each syllable. In step 4, we merge the Arabic script for each syllable and get the generated candidate loanword string. The multi-step process that takes place in step 2 of the process is:
Step 2.1
: Make minor vowel adjustments in certain contexts, e.g., when 'a' is between two consonants it is changed to 'A'.
Step 2.2
: Perform bulk of conversion by using table of mappings from IPA characters to modified Buckwalter characters such as 'a'\(\rightarrow\)'a','k'\(\rightarrow\)'k', 'y:'\(\rightarrow\)'iy', etc. that were supplied by a language expert.
Step 2.3
: Perform miscellaneous modifications to finalize the modified Buckwalter strings, e.g., if a syllable ends in 'a', then append an 'A' to that syllable.
The entire conversion process is illustrated in Figure [\[f:exampleConversion\]](#f:exampleConversion){reference-type="ref" reference="f:exampleConversion"} for the French word *raconteur*. At the top of the Figure is the IPA from the French dictionary entry with syllables marked. At the next level, step 1 (syllabification) has been completed. Step 2.1 doesn't apply to any of the syllables in this word since there are no minor vowel adjustments that are applicable for this word so at the next level each syllable is shown after step 2.2 has been completed. The next level shows the syllables after step 2.3 has been completed. The next level shows after step 3 has been completed and then at the end the strings are merged to form the candidate loanword.
# Experiments and Discussion {#experiments}
In our experiments we extracted a French-English bilingual dictionary using the freely available English Wiktionary dump 20131101 downloaded from <http://dumps.wikimedia.org/enwiktionary>. From this dump we extracted all the French words, their pronunciations, and their English definitions. Using the process described in section [3](#method){reference-type="ref" reference="method"} to convert each of the French pronunciations into Arabic script yielded 8277 unique loanword candidate strings. The data used for testing consists of a million lines of user comments crawled from the Moroccan news website <http://www.hespress.com>. The crawled user comments contain Moroccan Darija in heavily code-switched environments. While this makes for a challenging setting, it is a realistic representation of the types of environments in which historically unwritten languages are being written for the first time. The data we used is consistent with well-known code-switching among Arabic speakers, extending spoken discourse into formal writing. The total number of tokens in our Hespress corpus is 18,781,041. We found that 1150 of our 8277 loanword candidates appear in our Hespress corpus. Moreover, more than a million (1169087) loanword candidate instances appear in the corpus. Recall that a match could be a true match that really is a French loanword or a false match that just happens to coincidentally have string equality with words in the borrowing language, but is not a French loanword. False matches are particularly likely to occur for very short words. Accordingly, we filter out candidates that are of length less than four characters. This leaves us with 838 candidates appearing in the corpus and 217616 candidate instances in the corpus. To get an idea of what percentage of our matches are true matches versus false matches, we conducted an annotation exercise with two native Moroccan Darija speakers who also knew at least intermediate French. We pulled a random sample[^3] of 1185 candidate instances from our corpus and asked each annotator to mark each instance as either:
A
: if the instance is originally from Arabic,
F
: if the instance is originally from French, or
U
: if they were not sure.
The results are shown in Table [1](#t:annotationResults){reference-type="ref" reference="t:annotationResults"}. There are a substantial number of French loanwords that are found. Some examples of translations successfully induced by our method are:
omelette
: اومليت\>; and
bourgeoisie
: بورجوازي\>.
:::
We hypothesize that our method can help improve machine translation (MT) of historically unwritten dialects with nearly zero resources. To test this hypothesis, we ran an MT experiment as follows. First we selected a random set of sentences from the Hespress corpus that each contained at least one candidate instance and had an MSA/Moroccan Darija/English trilingual translator translate them into English. In total, 273 sentences were translated. This served as our test set. We trained a baseline MT system using all GALE MSA-English parallel corpora available from the Linguistic Data Consortium (LDC) from 2007 to 2013.[^4] We trained the system using Moses 3.0 with default parameters. This baseline system achieves BLEU score of 7.48 on our difficult test set of code-switched Moroccan Darija and MSA. We trained a second system using the parallel corpora with our induced Moroccan Darija-English translation lexicon appended to the end of the training data. This time the BLEU score increased to 8.11, a gain of.63 BLEU points.
# Conclusions
With the explosive growth of informal textual electronic communications such as social media, web comments, etc., many colloquial everyday languages that were historically unwritten are now being written for the first time often in heavily code-switched text with traditionally written languages. The new written versions of these languages pose significant challenges for multilingual processing technology due to Out-Of-Vocabulary (OOV) challenges. Yet it is relatively common that these historically unwritten languages borrow significant amounts of vocabulary from relatively well resourced written languages. We presented a method for translation lexicon induction via loanwords for alleviating the OOV challenges in these settings where the borrowing language has extremely limited amounts of resources available, in many cases not even substantial amounts of monolingual data that is typically exploited by previous cognates and loanword detection methods to induce translation lexicons. This paper demonstrates induction of a Moroccan Darija-English translation lexicon via bridging French loanwords using the method and in MT experiments, the addition of the induced Moroccan Darija-English lexicon increased system performance by.63 BLEU points. | {'timestamp': '2017-08-22T02:07:39', 'yymm': '1706', 'arxiv_id': '1706.01570', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01570'} |
# Introduction
Summarization of large texts is still an open problem in language processing. People nowadays have lesser time and patience to go through large pieces of text which make automatic summarization important. Automatic summarization has significant applications in summarizing large texts like stories, journal papers, news articles and even larger texts like books. Existing methods for summarization can be broadly categorized into two categories *Extractive* and *Abstractive*. Extractive methods picks up words and sometimes directly sentences from the text. These methods are inherently limited in the sense that they can never generate human level summaries for large and complicated documents which require rephrasing sentences and incorporating information from full text to generate summaries. Most of the work done on summarization in past has been extractive. On the other hand most Abstractive methods take advantages of the recent developments in deep learning. Specifically the recent success of the sequence to sequence learning models where recurrent networks read the text, encodes it and then generate target text. Though these methods have recently shown to be competitive with the extractive methods they are still far away from reaching human level quality in summary generation. The work on summarization using AMR was started by . Abstract Meaning Representation (AMR) was as introduced by . AMR focuses on capturing the meaning of the text, by giving a specific meaning representation to the text. AMR tries to capture the *\"who is doing what to whom\"* in a sentence. The formalism aims to give same representation to sentences which have the same underlying meaning. For example *\"He likes apple\"* and *\"Apples are liked by him\"* should be assigned the same AMR. 's approach aimed to produce a summary for a story by extracting a summary subgraph from the story graph and finally generate a summary from this extracted graph. But, because of the unavailability of AMR to text generator at that time their work was limited till extracting the summary graph. This method extracts a single summary graph from the story graph. Extracting a single summary graph assumes that all of the important information from the graph can be extracted from a single subgraph. But, it can be difficult in cases where the information is spread out in the graph. Thus, the method compromises between size of the summary sub-graph and the amount of information it can extract. This can be easily solved if instead of a single sub-graph, we extract multiple subgraphs each focusing on information in a different part of the story. We propose a two step process for extracting multiple summary graphs. First step is to select few sentences from the story. We use the idea that there are only few sentences that are important from the point of view of summary, i.e. most of the information contained in the summary is present in very few sentences and they can be used to generate the summary. Second step is to extract important information from the selected sentences by extracting a sub-graph from the selected sentences. Our main contributions in this work are three folds,
- We propose a full-fledged pipeline for text summarization, providing strong baseline for future work on summarization using AMR.
- Present a novel approach for extracting multiple summary graphs that outperforms the previous methods based on a single sub-graph extraction.
- Expose some problems with existing evaluation methods and datasets for abstractive summarization.
The rest of the paper is organized as follows. Section [2](#sec:2){reference-type="ref" reference="sec:2"} contains introduction to AMR, section [3](#sec:3){reference-type="ref" reference="sec:3"} and [4](#sec:4){reference-type="ref" reference="sec:4"} contains the datasets and the algorithm used for summary generation respectively. Section [5](#sec:5){reference-type="ref" reference="sec:5"} has a detailed step-by-step evaluation of the pipeline and in section [6](#sec:6){reference-type="ref" reference="sec:6"} we discuss the problems with the current dataset and evaluation metric.
# Background: AMR Parsing and Generation {#sec:2}
AMR was introduced by with the aim to induce work on statistical Natural Language Understanding and Generation. AMR represents meaning using graphs. AMR graphs are rooted, directed, edge and vertex labeled graphs. Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} shows the graphical representation of the AMR graph of the sentence *\"I looked carefully all around me\"* generated by JAMR parser ( ). The graphical representation was produced using AMRICA . The nodes in the AMR are labeled with *concepts* as in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} *around* represents a concept. Edges contains the information regarding the *relations* between the concepts. In Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} *direction* is the relation between the concepts *look-01* and *around*. AMR relies on Propbank for semantic relations (edge labels). Concepts can also be of the form *run-01* where the index *01* represents the first sense of the word *run*. Further details about the AMR can be found in the AMR guidelines . A lot of work has been done on parsing sentences to their AMR graphs. There are three main approaches to parsing. There is alignment based parsing (JAMR-Parser), which uses graph based algorithms for concept and relation identification. Second, grammar based parsers like (CAMR) generate output by performing shift reduce transformations on output of a dependency parser. Neural parsing is based on using *seq2seq* models for parsing, the main problem for neural methods is the absence of a huge corpus of human generated AMRs. reduced the vocabulary size to tackle this problem while used larger external corpus of external sentences. Recently, some work has been done on producing meaningful sentences form AMR graphs. used a number of tree to string conversion rules for generating sentences. reformed the problem as a traveling salesman problem. used *seq2seq* learning methods.
# Datasets {#sec:3}
We used two datasets for the task-AMR Bank and CNN-Dailymail ( ). We use the proxy report section of the AMR Bank, as it is the only one that is relevant for the task because it contains the gold-standard (human generated) AMR graphs for news articles, and the summaries. In the training set the stories and summaries contain 17.5 sentences and 1.5 sentences on an average respectively. The training and test sets contain 298 and 33 summary document pairs respectively. CNN-Dailymail corpus is better suited for summarization as the average summary size is around 3 or 4 sentences. This dataset has around 300k document summary pairs with stories having 39 sentences on average. The dataset comes in 2 versions, one is the anonymized version, which has been preprocessed to replace named entities, e.g., *The Times of India*, with a unique identifier for example *\@entity1*. Second is the non-anonymized which has the original text. We use the non-anonymized version of the dataset as it is more suitable for AMR parsing as most of the parsers have been trained on non-anonymized text. The dataset does not have gold-standard AMR graphs. We use automatic parsers to get the AMR graphs but they are not gold-standard and will effect the quality of final summary. To get an idea of the error introduced by using automatic parsers, we compare the results after using gold-standard and automatically generated AMR graphs on the gold-standard dataset.
# Pipeline for Summary Generation {#sec:4}
The pipeline consists of three steps, first convert all the given story sentences to there AMR graphs followed by extracting summary graphs from the story sentence graphs and finally generating sentences from these extracted summary graphs. In the following subsections we explain each of the methods in greater detail.
## Step 1: Story to AMR
As the first step we convert the story sentences to their Abstract Meaning Representations. We use JAMR-Parser version 2 as it's openly available and has a performance close to the state of the art parsers for parsing the CNN-Dailymail corpus. For the AMR-bank we have the gold-standard AMR parses but we still parse the input stories with JAMR-Parser to study the effect of using graphs produced by JAMR-Parser instead of the gold-standard AMR graphs.
## Step 2: Story AMR to Summary AMR
After parsing (Step 1) we have the AMR graphs for the story sentences. In this step we extract the AMR graphs of the summary sentences using story sentence AMRs. We divide this task in two parts. First is finding the important sentences from the story and then extracting the key information from those sentences using their AMR graphs.
### Selecting Important Sentences {#ssec:impsnt}
Our algorithm is based on the idea that only few sentences are important from the point of view of summary i.e. there are only a few sentences which contain most of the important information and from these sentences we can generate the summary. **Hypothesis:** Most of the information corresponding to a summary sentence can be found in only one sentence from the story. To test this hypothesis, for each summary sentence we find the sentence from the story that contains maximum information of this summary sentence. We use ROGUE-1 Recall scores (measures the ratio of number of words in the target summary that are contained in the predicted summary to the total number of words in the target summary) as the metric for the information contained in the story sentence. We consider the story sentence as the predicted summary and the summary sentence as the target summary. The results that we obtained for 5000 randomly chosen document summary pairs from the CNN-Dailymail corpus are given in figure [\[graph\]](#graph){reference-type="ref" reference="graph"}. The average recall score that we obtained is 79%. The score will be perfectly 1 when the summary sentence is directly picked up from a story sentence. Upon manual inspection of the summary sentence and the corresponding best sentence from the story we realized, when this score is more than 0.5 or 0.6, almost always the information in the summary sentence is contained in this chosen story sentence. The score for in these cases is not perfectly 1 because of stop words and different verb forms used in story and summary sentence. Around 80% of summary sentences have score above 0.5. So, our hypothesis seems to be correct for most of the summary sentences. This also suggests the highly extractive nature of the summary in the corpus. Now the task in hand is to select few important sentences. Methods that use sentence extraction for summary generation can be used for the task. It is very common in summarization tasks specifically in news articles that a lot of information is contained in the initial few sentences. Choosing initial few sentences as the summary produces very strong baselines which the state-of-the-art methods beat only marginally. Ex. On the CNN-Dailymail corpus the state-of-the-art extractive method beats initial 3 sentences only by 0.4% as reported by . Using this idea of picking important sentences from the beginning, we propose two methods, first is to simply pick initial few sentences, we call this *first-n* method where n stands for the number of sentences. We pick initial 3 sentences for the CNN-Dailymail corpus i.e. *first-3* and only the first sentence for the proxy report section (AMR Bank) i.e. *first-1* as they produce the best scores on the ROGUE metric compared to any other *first-n*. Second, we try to capture the relation between the two most important entities (we define importance by the number of occurrences of the entity in the story) of the document. For this we simply find the first sentence which contains both these entities. We call this the *first co-occurrence* based sentence selection. We also select the first sentence along with *first co-occurrence* based sentence selection as the important sentences. We call this the *first co-occurrence+first* based sentence selection.
### Extracting Summary Graph {#ssec:extract-AMR}
As the datasets under consideration are news articles. The most important information in them is about an entity and a verb associated with it. So, to extract important information from the sentence. We try to find the entity being talked about in the sentence, we consider the most referred entity (one that occurs most frequently in the text), now for the main verb associated with the entity in the sentence, we find the verb closest to this entity in the AMR graph. We define the closest verb as the one which lies first in the path from the entity to the root. We start by finding the position of the most referred entity in the graph, then we find the closest verb to the entity. and finally select the subtree hanging from that verb as the summary AMR.
## Step 3: Summary Generation
To generate sentences from the extracted AMR graphs we can use already available generators. We use Neural AMR ( ) as it provides state of the art results in sentence generation. We also use (JAMR-Generator) in one of the experiments in the next section. Generators significantly effect the results, we will analyze the effectiveness of generator in the next section.
# Results and Analysis {#sec:5}
## Baselines {#ssec:baselines}
[\[ssec:evaluation\]]{#ssec:evaluation label="ssec:evaluation"}
In this section, we present the baseline models and analysis method used for each step of our pipeline. For the CNN-Dailymail dataset, the Lead-3 model is considered a strong baseline; both the abstractive and extractive state-of-the art methods on this dataset beat this baseline only marginally. The Lead-3 model simply produces the leading three sentences of the document as its summary. The key step in our pipeline is step-2 i.e. summary graph extraction. Directly comparing the Lead-3 baseline, with AMR based pipeline to evaluate the effectiveness of step-2 is an unfair comparison because of the errors introduced by imperfect parser and generator in the AMR pipeline. Thus to evaluate the effectiveness of step-2 against Lead-3 baseline, we need to nullify the effect of errors introduce by AMR parser and generator. We achieve this by trying to introduce similar errors in the leading thre sentences of each document. We generate the AMR graphs of the leading three sentences and then generate the sentences using these AMR graph. We use parser and generator that were used in our pipeline. We consider these generated sentences as the new baseline summary, we shall now refer to it *Lead-3-AMR* baseline in the remaining of the paper. For the proxy report section of the AMR bank, we consider the *Lead-1-AMR* model as the baseline. For this dataset we already have the gold-standard AMR graphs of the sentences. Therefore, we only need to nullify the error introduced by the generator.
## Procedure to Analyze and Evaluate each step
For the evaluation of summaries we use the standard ROGUE metric. For comparison with previous AMR based summarization methods, we report the *Recall*, *Precision* and *\(F_{1}\)* scores for ROGUE-1. Since most of the literature on summarization uses *\(F_{1}\)* scores for ROGUE-2 and ROGUE-L for comparison, we also report *\(F_{1}\)* scores for ROGUE-2 and ROGUE-L for our method. ROGUE-1 Recall and Precision are measured for uni-gram overlap between the reference and the predicted summary. On the other hand, ROGUE-2 uses bi-gram overlap while ROGUE-L uses the longest common sequence between the target and the predicted summaries for evaluation. In rest of this section, we provide methods to analyze and evaluate our pipeline at each step. **Step-1: AMR parsing** To understand the effects of using an AMR parser on the results, we compare the final scores after the following two cases-first, when we use the gold-standard AMR graphs and second when we used the AMR graphs generated by JAMR-Parser in the pipeline. Section [5.4](#ssec:jamr_effect){reference-type="ref" reference="ssec:jamr_effect"} contains a comparison between the two. **Step-2: Summary graph extraction** For evaluating the effectiveness of the summary graph extraction step we compare the final scores with the *Lead-n-AMR* baselines described in section [5.1](#ssec:baselines){reference-type="ref" reference="ssec:baselines"}. In order to compare our summary graph extraction step with the previous work ( ), we generate the final summary using the same generation method as used by them. Their method uses a simple module based on alignments for generating summary after step-2. The alignments simply map the words in the original sentence with the node or edge in the AMR graph. To generate the summary we find the words aligned with the sentence in the selected graph and output them in no particular order as the predicted summary. Though this does not generate grammatically correct sentences, we can still use the ROGUE-1 metric similar to , as it is based on comparing uni-grams between the target and predicted summaries. **Step-3: Generation** For evaluating the quality of the sentences generated by our method, we compare the summaries generated by the first-1 model and Lead-1-AMR model on the gold-standard dataset. However, when we looked at the scores given by ROGUE, we decided to do get the above summaries evaluated by humans. This produced interesting results which are given in more detail in section [5.5](#ssec:generator_effect){reference-type="ref" reference="ssec:generator_effect"}.
## Results on the Proxy report section
In table [\[proxy-section-results\]](#proxy-section-results){reference-type="ref" reference="proxy-section-results"} we report the results of using the pipeline with generation using the alignment based generation module defined in section [\[ssec:evaluation\]](#ssec:evaluation){reference-type="ref" reference="ssec:evaluation"}, on the proxy report section of the AMR Bank. All of our methods out-perform 's method. We obtain best ROGUE-1 \(F_{1}\) scores using the *first co-occurrence+first* model for important sentences. This also out-perform our *Lead-1-AMR* baseline by 0.3 ROGUE-1 \(F_{1}\) points.
## Effects of using JAMR Parser {#ssec:jamr_effect}
In this subsection we analyze the effect of using JAMR parser for step-1 instead of the gold-standard AMR graphs. First part of table [\[final-results\]](#final-results){reference-type="ref" reference="final-results"} has scores after using the gold-standard AMR graphs. In the second part of table [\[final-results\]](#final-results){reference-type="ref" reference="final-results"} we have included the scores of using the JAMR parser for AMR graph generation. We have used the same Neural AMR for sentence generation in all methods. Scores of all methods including the *Lead-1-AMR* baseline have dropped significantly. The usage of JAMR Parser has affected the scores of *first co-occurrence+first* and *first-1* more than that for the *Lead-1-AMR*. The drop in ROGUE \(F_{1}\) score when we use *first co-occurrence+first* is around two ROGUE \(F_{1}\) points more than when *Lead-1-AMR*. This is a surprising result, and we believe that it is worthy of further research.
## Effectiveness of the Generator {#ssec:generator_effect}
In this subsection we evaluate the effectiveness of the sentence generation step. For fair comparison at the generation step we use the gold-standard AMRs and don't perform any extraction in step-2 instead we use full AMRs, this allows to remove any errors that might have been generated in step-1 and step-2. In order to compare the quality of sentences generated by the AMR, we need a gold-standard for sentence generation step. For this, we simply use the original sentence as gold-standard for sentence generation. Thus, we compare the quality of summary generated by *Lead-1* and *Lead-1-AMR*. The scores using the ROGUE metric are given in bottom two rows of table [\[human_evaluation\]](#human_evaluation){reference-type="ref" reference="human_evaluation"}. The results show that there is significant drop in *Lead-1-AMR* when compared to *Lead-1*. We perform human evaluation to check whether the drop in ROGUE scores is because of drop in information contained, and human readability or is it because of the inability of the ROGUE metric to judge. To perform this evaluation we randomly select ten test examples from the thirty-three test cases of the proxy report section. For each example, we show the summaries generated by four different models side by side to the human evaluators. The human evaluator does not know which summaries come from which model. A score from 1 to 10 is then assigned to each summary on the basis of *readability*, and *information contained* of summary, where 1 corresponds to the lower level and 10 to the highest. In table [\[human_evaluation\]](#human_evaluation){reference-type="ref" reference="human_evaluation"} we compare the scores of these four cases as given by ROGUE along with human evaluation. The parser-generator pairs for the four cases are gold-JAMR(generator), JAMR(parser)-neural, gold-neural, and the original sentence respectively. Here gold parser means that we have used the gold-standard AMR graphs. The scores given by the humans do not correlate with ROGUE. Human evaluators gives almost similarly scores to summary generated by the *Lead-1* and *Lead-1-AMR* with *Lead-1-AMR* actually performing better on *readability* though it dropped some information as clear from the scores on *information contained*. On the other hand, ROGUE gives very high score to Lead-1 while models 1,2 and 4 get almost same scores. The similar scores of model 2 and 3 shows that generators are actually producing meaningful sentences. Thus the drop in ROGUE scores is mainly due to the inability of the ROGUE to evaluate abstractive summaries. Moreover, the ROGUE gives model 4 higher score compared to model 1 while human evaluators give the opposite scores on *information contained* in the sentence. A possible reason for the inability of the ROGUE metric to properly evaluate the summaries generated by our method might be due to its inability to evaluate restructured sentences. AMR formalism tries to assign the same AMR graphs to the sentences that have same meaning so there exists a one-to-many mapping from AMR graphs to sentences. This means that the automatic generators that we are using might not be trying to generate the original sentence; instead it is trying to generate some other sentence that has the same underlying meaning. This also helps in explaining the low ROGUE-2 and ROGUE-L scores. If the sentences might be getting rephrased, they would loose most of the bi-and tri-grams from the original sentence resulting in low ROGUE-2 and ROGUE-L scores.
## Analyzing the effectiveness of AMR extraction
The aim of extracting summary graphs from the AMR graphs of the sentence is to drop the not so important information from the sentences. If we are able to achieve this perfectly, the ROGUE-1 Recall scores that we are getting should remain almost the same (since we are not add any new information) and the ROGUE-1 precision should go up (as we have thrown out some useless information); thus effectively improving the overall ROGUE-1 \(F_{1}\) score. In the first two rows of table [\[final-results\]](#final-results){reference-type="ref" reference="final-results"} we have the scores after using the full-AMR and extracted AMR for generation respectively. It is safe to say that extracting the AMR results in improved ROGUE-1 precision whereas ROGUE-1 Recall reduces only slightly, resulting in an overall improved ROGUE-1 \(F_{1}\).
## Results on the CNN-Dailymail corpus
In table [\[fig:CNN-Results\]](#fig:CNN-Results){reference-type="ref" reference="fig:CNN-Results"} we report the results on the CNN-Dailymail corpus. We present scores by using the *first-3* model. The first row contains the *Lead-3-AMR* baseline. The results we achieve are competitive with the *Lead-3-AMR* baseline. The rest of the table contains scores of Lead-3 baseline followed by the state-of-the-art method on the anonymized and non-anonymized versions of the dataset. The drop in the scores from the Lead-3(non-anonymized) to *Lead-3-AMR* is significant and is largely because of the error introduced by parser and generator.
# Related Work and Discussion {#sec:6}
## Related Work
showed that most of the work in text summarization has been extractive, where sentences are selected from the text which are then concatenated to form a summary. transformed the input to nodes, then used the Pagerank algorithm to score nodes, and finally grow the nodes from high-value to low-value using some heuristics. Some of the approaches combine this with sentence compression, so more sentences can be packed in the summary. , , , and among others used ILPs and approximations for encoding compression and extraction. Recently some abstractive approaches have also been proposed most of which used sequence to sequence learning models for the task. , , , used standard encoder-decoder models along with their variants to generate summaries. incorporated the AMR information in the standard encoder-decoder models to improve results. Our work in similar to other graph based abstractive summarization methods and . used dependency parse trees to produce summaries. On the other hand our work takes advantage of semantic graphs.
## Need of an new Dataset and Evaluation Metric {#ssec:rogue_problems}
ROGUE metric, by it is design has lots of properties that make it unsuitable for evaluating abstractive summaries. For example, ROGUE matches exact words and not the stems of the words, it also considers stop words for evaluation. One of the reasons why ROGUE like metrics might never become suitable for evaluating abstractive summaries is its incapabilities of knowing if the sentences have been restructured. A good evaluation metric should be one where we compare the meaning of the sentence and not the exact words. As we showed section [5.5](#ssec:generator_effect){reference-type="ref" reference="ssec:generator_effect"} ROGUE is not suitable for evaluating summaries generated by the AMR pipeline. We now show why the CNN-Dailymail corpus is not suitable for Abstractive summarization. The nature of summary points in the corpus is highly extractive (Section [4.2.1](#ssec:impsnt){reference-type="ref" reference="ssec:impsnt"} for details) where most of the summary points are simply picked up from some sentences in the story. Tough, this is a good enough reason to start searching for better dataset, it is not the biggest problem with the dataset. The dataset has the property that a lot of important information is in the first few sentences and most of the summary points are directly pick from these sentences. The extractive methods based on sentence selection like SummaRunNer are not actually performing well, the results they have got are only slightly better than the Lead-3 baseline. The work doesn't show how much of the selected sentences are among the first few and it might be the case that the sentences selected by the extractive methods are mostly among the first few sentences, the same can be the problem with the abstractive methods, where most the output might be getting copied from the initial few sentences. These problems with this corpus evoke the need to have another corpus where we don't have so much concentration of important information at any location but rather the information is more spread out and the summaries are more abstractive in nature.
# Possible Future Directions
As this proposed algorithm is a step by step process we can focus on improving each step to produce better results. The most exciting improvements can be done in the summary graph extraction method. Not a lot of work has been done to extract AMR graphs for summaries. In order to make this pipeline generalizable for any sort of text, we need to get rid of the hypothesis that the summary is being extracted exactly from one sentence. So, the natural direction seems to be joining AMR graphs of multiple sentences that are similar and then extracting the summary AMR from that large graph. It will be like clustering similar sentences and then extracting a summary graph from each of these cluster. Another idea is to use AMR graphs for important sentence selection.
# Conclusion
In this work we have explored a full-fledged pipeline using AMR for summarization for the first time. We propose a new method for extracting summary graph, which outperformed previous methods. Overall we provide strong baseline for text summarization using AMR for possible future works. We also showed that ROGUE can't be used for evaluating the abstractive summaries generated by our AMR pipeline. | {'timestamp': '2017-07-19T02:02:19', 'yymm': '1706', 'arxiv_id': '1706.01678', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01678'} |
null | null |
# Introduction
In the standard cosmological model, the accelerated expansion of the universe is attributed to the cosmological constant \(\Lambda\). However, to match the observed expansion, \(\Lambda\) must be of the order of \(10^{-122}\) in Planck units, which raises a fine-tuning problem. A possible alternative is to modify general relativity (GR) at large distances or low momenta. A massive spin-2 field theory, known as the dRGT theory , is a theoretically well-motivated modification of GR. However, the dRGT theory does not admit a flat Friedmann-Lemaı̂tre-Robertson-Walker (FLRW) solution with an expanding scale factor . A modification to dRGT gravity, known as Quasi-Dilaton Massive Gravity (QDMG), was proposed in and provides homogeneous and isotropic expanding solutions. It was later shown that the parameters of QDMG have to be finely tuned in order to match the observed expansion history of the universe. More disastrously, the results in indicate that the scalar perturbations in QDMG acquire a wrong sign kinetic term at short scales. A further modification, Extended QDMG (EQDMG), was proposed in and has scalar perturbations that are thought to be stable at all momentum scales. The standard fixed-point cosmological solution of EQDMG has a de Sitter metric, and thus appears to be a good candidate for late-time cosmology. EQDMG only differs from QDMG by the addition to the action of one operator involving the Quasi-Dilaton (QD) field and a new free parameter \(\alpha_{\sigma}\). Naively, in the limit \(\alpha_{\sigma}\to 0\), EQDMG reduces to QDMG, but actually the limit is very subtle. Indeed, there are controversial results in the literature regarding whether or not EQDMG contains an unavoidable additional degree of freedom, which would correspond to the Boulware-Deser ghost (BD). In this paper, for the first time, we:
1. assess the stability of the standard fixed-point solutions (referred to as Case 1 in this paper) and show that this assessment requires a non-standard approach;
2. demonstrate that the Case 1 fixed points cannot be approached dynamically, due to an unavoidable singularity in the dynamical equations;
3. perform a comprehensive study of a new branch of solutions (referred to as Case 2), first proposed by but largely ignored in the literature, and show that it provides stable and dynamically attainable fixed-point solutions;
4. show for a flat FLRW universe, the fact that the background equations are satisfied does not guarantee the presence of the additional constraint necessary to eliminate the BD mode (in agreement with the results of, but in disagreement with the computations in );
5. find that the only branch of solutions for which the additional constraint exists corresponds to Case 1;
6. argue that, in order to avoid a BD ghost, the initial values of certain EQDMG dynamical variables must be extremely fine-tuned;
7. verify that the same fine-tuned initial conditions also allow the fixed point to be approached dynamically.
The paper is organized as follows. In Section [2](#Formalism){reference-type="ref" reference="Formalism"} and [3](#Background){reference-type="ref" reference="Background"}, we summarize the theory of EQDMG. In Section [4](#DynEqnsVar){reference-type="ref" reference="DynEqnsVar"}, we define the dynamical variables and provide the relevant background equations. In Section [5](#fixedpoint){reference-type="ref" reference="fixedpoint"}, we find the fixed-point and de Sitter solutions of the dynamical equations, and show that they are equivalent to one another (provided the Hubble rate is positive). We identify four independent fixed-point cases, each of which is studied in greater detail in the sections that follow. In Section [6](#stability){reference-type="ref" reference="stability"}, we introduce the standard procedure for analyzing the stability of the fixed-point solutions for the background. We discuss the inadequacy of this procedure for the Case 1 fixed points and provide an augmented framework. In Section [7](#scanning){reference-type="ref" reference="scanning"}, we present the results of our numerical search for viable parameters for the EQDMG theory. We find that except for a very specific, precisely fine-tuned initial displacement away from the fixed-point values of Case 1 (explained in Section [8](#bdghost){reference-type="ref" reference="bdghost"}), the fixed points cannot be reached in the asymptotic future. In Section [8](#bdghost){reference-type="ref" reference="bdghost"}, we further study the perturbative stability of the scalar sector of the theory, both in the vacuum case and with matter. We identify the conditions on the dynamical variables required to avoid the BD ghost, and show that the fact the background equations are satisfied does not guarantee the validity of the conditions. Cases 1 turns out to be the only case for which the additional constraint necessary to eliminate the BD mode can be obtained by an appropriate choice of initial conditions. However, those conditions appear to represent a difficult fine-tuning of all the degrees of freedom. In Section [9](#cnclsns){reference-type="ref" reference="cnclsns"} we present our conclusions. We provide some detailed calculations and consider the special case of Minkowski solutions in the appendices.
# Formalism {#Formalism}
We consider the action for the extended quasi-dilaton theory: \[\begin{aligned}
S & = & S_{EH}+S_{m}+S_{\sigma} = \frac{M_{{\rm Pl}}^{2}}{2}\int d^{4}x\sqrt{-g}\biggl[R\\ \nonumber &-&\frac{\omega}{M_{{\rm Pl}}^{2}}\partial_{\mu}\sigma\partial^{\mu}\sigma+2m_{g}^{2}(\mathcal{L}_{2}+\alpha_{3}\mathcal{L}_{3}+\alpha_{4}\mathcal{L}_{4})\biggr]+S_{m},\label{eq:action} \end{aligned}\] where \(M_{\rm Pl}\) is the Planck mass and, in addition to the Einstein Hilbert action \(S_{EH}\), we have the contribution \(S_m\) of the matter sector, and a quasi-dilaton contribution \(S_{\sigma}\). Here \[\begin{aligned}
\mathcal{L}_{2} & \equiv & \frac{1}{2}\,([\mathcal{K}]^{2}-[\mathcal{K}^{2}])\,,\\ \mathcal{L}_{3} & \equiv & \frac{1}{6}\,([\mathcal{K}]^{3}-3[\mathcal{K}][\mathcal{K}^{2}]+2[\mathcal{K}^{3}])\,,\\ \mathcal{L}_{4} & \equiv & \frac{1}{24}\,([\mathcal{K}]^{4}-6[\mathcal{K}]^{2}[\mathcal{K}^{2}]+3[\mathcal{K}^{2}]^{2}\\ \nonumber &+&8[\mathcal{K}][\mathcal{K}^{3}]-6[\mathcal{K}^{4}])\,, \end{aligned}\] with square brackets denoting a trace, and \[\begin{aligned}
\label{ktensor} \mathcal{K}_{\ \nu}^{\mu}&=&\delta_{\ \nu}^{\mu}-e^{\sigma/M_{{\rm Pl}}}\left(\sqrt{g^{-1}{f}}\right)_{\ \ \nu}^{\mu}\,,\\ {f}_{\mu\nu}&\equiv&\eta_{ab}\partial_{\mu}\phi^{a}\partial_{\nu}\phi^{b}-\frac{\alpha_{\sigma}}{M_{{\rm Pl}}^{2}m_{g}^{2}}e^{-2\sigma/M_{{\rm Pl}}}\partial_{\mu}\sigma\partial_{\nu}\sigma. \end{aligned}\] \(S_{\sigma}\) includes five new fields: \(\sigma\) is the quasi-dilaton scalar field and \(\phi^{a}\) (\(a=0,\cdots,3\)) are the four Stückelberg fields. It also depends on the coupling constants \(\alpha_{\sigma}\), \(\alpha_{2}\) and \(\alpha_{3}\), and on the graviton mass \(m_{g}\). For \(\alpha_{\sigma}=0\) one recovers the standard quasi-dilaton theory. In the space of Stückelberg fields, the theory enjoys the Poincare symmetry \[\phi^{a}\to\phi^{a}+c^{a}\,,\qquad\phi^{a}\to\Lambda_{b}^{a}\phi^{b}\,,\] and a global symmetry given by \[\sigma\to\sigma+\sigma_{0}\,,\qquad\phi^{a}\to e^{-\sigma_{0}/M_{{\rm Pl}}}\,\phi^{a}\,,\label{eqn:symmetry}\] with \(\sigma_{0}\) an arbitrary constant.
# Background {#Background}
We consider a spatially flat Friedmann-Lemaı̂tre-Robertson-Walker (FLRW) ansatz, for which \[\begin{aligned}
ds^{2} & = &-\mathcal{N}(t)^{2}dt^{2}+a(t)^{2}\delta_{ij}dx^{i}dx^{j}\,,\\ \phi^{0} & = & \phi^{0}(t)\,,\\ \phi^{i} & = & x^{i}\,,\\ \sigma & = & \bar{\sigma}(t)\. \end{aligned}\] The fiducial metric \(f_{\mu\nu}\) reduces to \[{f}_{00}=-n(t)^2\,, \quad {f}_{ij}=\delta_{ij}\,,\] where \[\label{defn} n(t)^2 \equiv \bigl(\dot{\phi}^{0}\bigr)^{2} +\frac{\alpha_{\sigma}}{M_{\rm Pl}^2m_g^2}\,e^{-2\bar{\sigma}/M_{\rm Pl}} {\dot{\bar{\sigma}}}^2\.\] The minisuperspace action for the background can be written as \[\begin{aligned}
S/V & = & M_{{\rm Pl}}^{2}\int dt\biggl[-3 \frac{a^3}{\mathcal{N}}\left(\frac{\dot{a}}{a}\right)^2+a^3\frac{w}{M_{\rm Pl}}\frac{\dot{\sigma}^2}{2\mathcal{N}}\\ \nonumber & + &\mathcal{N} a^3 m_{g}^{2}(\mathcal{L}_{2}+\alpha_{3}\mathcal{L}_{3}+\alpha_{4}\mathcal{L}_{4})\biggr], \label{eq:actionB} \end{aligned}\] where \[\begin{aligned}
\mathcal{L}_{2} &= &3(X-1)(-2+X(1+r))\,,\\ \mathcal{L}_{3} & =&-(X-1)^2(-4+X(1+3r))\,,\\ \mathcal{L}_{4} & = &(X-1)^3(-1+rX)\,, \end{aligned}\] and we have defined \[\begin{aligned}
\label{xsigma} X & \equiv & \frac{e^{\bar{\sigma}/M_{\rm Pl}}}{a}\,,\\ r & \equiv & \frac{n}{\mathcal{N}}\, a\,\label{rna}. \end{aligned}\] It is worth pointing out here that in , \[\left(\sqrt{g^{-1}f}\right)^{\mu}_{\nu}= \begin{bmatrix} \pm{\nicefrac{n}{\mathcal{N}}} & 0 & 0 & 0 \\ 0 & \pm{\nicefrac{1}{a}} & 0 & 0\\ 0 & 0 & \pm{\nicefrac{1}{a}} & 0 \\ 0 & 0 & 0 & \pm{\nicefrac{1}{a}} \end{bmatrix}\] and we make the \((+,+,+,+)\) choice following . This gives us \[r>0.\] \(r=0\) represents a determinant singularity in either or both \(f_{\mu\nu}\) (if \(n=0\)) or \(g_{\mu\nu}\) (if \(a=0\))--a spacelike hypersurface where the dimensionality of the metric changes. The stability of the theory across that hypersurface is unclear . Indeed, we find that when \(r\) approaches too close to zero, our numerical integrations of the dynamical equations become unstable; the instability in the numerical noise may be due to an underlying instability in the theory. We insist here that \(r\neq0\).
# Dynamical Equations and Variables {#DynEqnsVar}
We next set out the dynamical variables describing the background, and the dynamical equations describing their time evolution. Varying the action with respect to \(\phi^0(t)\) leads to the constraint equation \[\label{Constraint} \partial_t\left[ \frac{\dot{\phi^0}}{n}a^4\, G_2(X)\right] = 0\,,\] where \(G_2(X)=X(1-X)J(X)\), with \[\label{jterm} J(X) \equiv 3+3(1-X)\alpha_{3}+(1-X)^{2}\alpha_{4}\.\] This suggests that it will be useful to introduce as one of the dynamical variables \[\label{ydef} y(t)\equiv\frac{\dot{\phi^0}(t) }{n}G_{2}(X) \.\] The solution of ([\[Constraint\]](#Constraint){reference-type="ref" reference="Constraint"}) is immediately \[\begin{aligned}
\label{yc} y&=&\frac{C}{a^4}\. \end{aligned}\] We see that, in any reasonable cosmological context, \(y\to0\) in the asymptotic future, and that this can be achieved by one (or more) of four quantities approaching or equalling zero: \(J(X(t))\), \(\dot{\phi^0}(t)\), \(X(t)\), or \(X(t)-1\). These four cases will drive our analysis. Anticipating that it will be convenient to regard \(X\) as a dynamical variable, we differentiate with respect to time to get \[\label{sigdot} \dot{X}=X\left(\frac{\dot{\sigma}}{M_{Pl}}-H\right)\.\] \(H(t)\equiv{\dot a}/a\) is the Hubble parameter. Varying the action with respect to the lapse \(\mathcal{N}(t)\) and using time reparametrization invariance to set \(\mathcal{N}(t)=1\), we obtain the Friedmann equation: \[\label{Friedman} 3H^2=\frac{\omega}{2}\left(\frac{\dot\sigma}{M_{\rm Pl}}\right)^2 + 3m_g^2 G_1(X)+\frac{\rho_m}{M_{\rm Pl}^2}+\frac{\rho_r}{M_{\rm Pl}^2}\,,\] where \[\label{G1} G_1(X)\equiv\frac{1}{3} (X-1) \left(J(X) + (X-1) (\alpha_3 (X-1)-3)\right)\.\] The form of the Friedman equation suggests regarding the first two terms on the right-hand side (RHS) of ([\[Friedman\]](#Friedman){reference-type="ref" reference="Friedman"}) as the dark energy density (divided by \(M_{\rm Pl}^2\)). For future convenience, we represent them separately as \[\label{omegaL} \Omega_{\Lambda}\equiv\frac{m_g^2}{H^2}G_1(X)\\ \] and \[\label{omegaS} \Omega_{\sigma}\equiv\frac{\omega}{6H^2}\left(\frac{\dot{\sigma}}{M_{\rm Pl}}\right)^2\,,\] and define \(\Omega_{DE}\equiv \Omega_{\Lambda}+\Omega_{\sigma}\). From the conservation of the stress-energy tensor of matter and of radiation we get, \[\begin{aligned}
\dot{\rho_m}&=&-3H{\rho_m},\label{omegam}\\ \dot{\rho_r}&=&-4H{\rho_r}\label{omegar}\. \end{aligned}\] From the conservation of the stress-energy tensor[^1] obtained from \(S_{\sigma}\), we get \[\label{conservSigma} (\ddot{\sigma}+3H\dot{\sigma})\omega\dot{\sigma} +3 M_{\rm Pl} m_g^2(\dot{\sigma}-r H M_{\rm Pl})X G_1'(X)=0,\] where a prime here stands for a derivative with respect to the argument \(X\) of the function. We select \(y,X,\tilde{\Omega}_i=\Omega_i {\tilde h}^2 (i=DE,m,r)\) as our dynamical variables, where \({\tilde h}\equiv H/m_g\) is the dimensionless Hubble parameter. We can now express the background evolution equations in terms of \(N\equiv log(a)\) (giving us \(dN=H dt\) assuming \(H \neq 0\))[^2]:
- From ([\[Constraint\]](#Constraint){reference-type="ref" reference="Constraint"}), we have immediately \[\frac{dy}{dN}=-4y \label{yN}\.\]
- The equations for \(\tilde{\Omega}_{m}\) and \(\tilde{\Omega}_{r}\) are similarly easily obtained from and : \[\begin{aligned}
\frac{ d\tilde{\Omega}_{m}}{dN}&=&-3 \tilde{\Omega}_{m},\label{matter}\\ \frac{ d\tilde{\Omega}_{r}}{dN}&=&-4\tilde{\Omega}_{r}\label{radiation}\. \end{aligned}\]
- Equation can be rewritten using , and the definition of \(\tilde{\Omega}_{DE}\): \[\frac{d X}{dN}= X\left(\pm\sqrt{\frac{6\left(\tilde{\Omega}_{DE}-G_{1}(X)\right)}{{\tilde h}^2\omega}} -1\right)\label{txx}\.\] The \(\pm\) represents the possibility that \(\dot\sigma\) can be positive or negative[^3].
- Equation can be rewritten using and : \[\label{tildeOmsigmaeqs} \frac{ d\tilde{\Omega}_{DE}}{dN} =-6\left(\tilde{\Omega}_{DE}-G_{1}(X)\right) +X G_1'(X) \left(r-1\right)\.\]
In the above set of equations one must replace
- \({\tilde h}^2\) by the Friedmann equation, which now takes the simple form \[\begin{aligned}
\label{hsqnew} {\tilde h}^2&=&\tilde{\Omega}_{DE}+\tilde{\Omega}_{m}+\tilde{\Omega}_{r}\,, \end{aligned}\]
- and (combining , and ) \(r\) with[^4] \[\begin{aligned}
\label{rdef} r&=&+\sqrt{\frac{6\left(\tilde{\Omega}_{DE}-G_{1}(X)\right)\alpha_{\sigma}}{\omega X^{2}\left(1-\left(\frac{y}{G_2(X)}\right)^2\right)}}. \end{aligned}\]
The argument of the square root on the right-hand side of must be positive for \(r\) to be real. *The reality of \(r\) is a condition on the dynamical variables that must be checked, in case (as we find below) it is not automatically satisfied.* In particular, we see that problems may arise if \(\left(y/G_2(X)\right)^2\to1\).
# Fixed-point Analysis {#fixedpoint}
In this section, we evaluate the dynamical variables when their \(N-\)derivatives vanish in -. We term the values of the dynamical variables in this limit as fixed points. In the fixed-point limit, the left-hand sides of equations - vanish, giving us \(y_{FP}=0\), and also \(\tilde{\Omega}_{m,FP}=\tilde{\Omega}_{r,FP}=0\). From , we learn that \(\tilde{\Omega}_{DE,FP}=\tilde{h}_{FP}^2\). The solutions to - are \[\begin{aligned}
y=y_0e^{-4N},\hskip 2pt \tilde{\Omega}_{m}=\tilde{\Omega}_{m0}e^{-3N}, \hskip 2pt\tilde{\Omega}_{r}=\tilde{\Omega}_{r0}e^{-4N}, \end{aligned}\] where \(y_0\), \(\tilde{\Omega}_{m0}\) and \(\tilde{\Omega}_{r0}\) are the corresponding initial values. Thus fixed points occur in the asymptotic future, i.e. as \(N\to\infty\) and so \(a \to \infty\). The left-hand side of vanishes at the fixed point, implying that \(X_{FP}\) is a constant. If \(X_{FP}\neq0\), the right-hand side of (and if \(X_{FP}=0\), then the right-hand side of ) provides us with \[\label{hsqfp} \tilde{\Omega}_{DE,FP}=\tilde{h}_{FP}^2 = \begin{cases} \frac{G_{1}(X_{FP})}{\left(1-\frac{\omega}{6}\right)}, & X_{FP}\neq0\,,\\ G_1(0), & X_{FP}=0\,, \end{cases}\] at the fixed point. In arriving at , we take the \(+\) sign in , since the \(-\) sign leads to \(X=0\) as the only fixed-point solution. Notice that, as for getting background fixed-point solutions, both \(0<\omega<6\) and \(\omega\geq6\) are in principle suitable regions in the parameter space, since \(G_{1}(X_{FP})\) can be either positive or negative. A special value is \(\omega=6\), in which case and demand \(G_{1}(X_{FP})=0\). Observing that \({\tilde h}^2\) is also a constant at the fixed point, we conclude that *the fixed points of the evolution equations are de Sitter*. We find that the converse is also true: *the de Sitter solutions of the evolution equations are fixed points as we approach the asymptotic future.* To prove this, we require that the dynamical variables attain the following de Sitter values in the future, \[\begin{aligned}
\left(\tilde{\Omega}_{DE},\tilde{\Omega}_m,\tilde{\Omega}_r\right) &=&\left(\tilde{h}_{FP}^2,0,0\right)\label{fpvals}, \end{aligned}\] where \(\tilde{h}_{FP}^2\) is a constant different from zero. In this situation, the left-hand sides of , and become zero, meaning they are fixed points. From , we learn that \(y = 0\) in the asymptotic future, which means that the left-hand side of is also zero. The only point left to establish is that \(X\) approaches a constant in the future. From the definitions , and , one can split the fixed-point solution into four cases:
- Case 1 (the standard case): \[\label{case1 } J(X_{FP})=0\,,\] and hence the fixed-point solutions are \[\label{xpm} X_{FP}=X_{\pm} = 1+\frac{3}{2}\frac{\alpha_3}{\alpha_4}\pm{\sqrt{\frac{9\alpha_{3}^{2}}{4\alpha_{4}^2}-\frac{3}{\alpha_4}}}.\]
- Case 2: \[\label{case2 } \left[\frac{\dot{\phi}^0}{n}\right]_{FP} = \sqrt{1-\alpha_{\sigma}\frac{6\left(\tilde{\Omega}_{DE,FP}-G_{1}(X_{FP})\right)}{\omega r^2_{FP} X^2_{FP}}}=0 \.\] Since the left-hand side of vanishes, provides us with the following equation in \(X_{FP}\) \[\begin{aligned}
\label{7th} &6(\tilde{h}^2_{FP}-G_{1}(X_{FP}))+X_{FP}G'_{1}(X_{FP})\\ \nonumber &=G'_{1}(X_{FP})\sqrt{\frac{6}{\omega }\alpha_{\sigma} \left(\tilde{h}^2_{FP}-G_{1}(X_{FP})\right)}. \end{aligned}\] Squaring gives us a polynomial equation for \(X_{FP}\). \(X_{FP}\) can be any of the roots of that polynomial.
- Case 3: \[\label{case3} X_{FP}=0\.\]
- Case 4: \[\label{case4} X_{FP}=1\.\]
For all cases, \(X\) approaches a constant and thus the left-hand side of vanishes in the asymptotic future proving it is a fixed point. We analyze the fixed-point solutions in more detail below.
## Case 1: \(J(X_{FP})=0\)
As will become clear below, this case is very subtle. Note that these fixed points are the same as ones analyzed in , for the QD theory (i.e., the EQD with \(\alpha_{\sigma}=0\)). Requiring that \(X_{\pm}\) be real, means[^5] \[\begin{aligned}
\alpha_{4}\leq\frac{3\alpha_{3}^{2}}{4}\. \end{aligned}\] From and , we get the same expression as in the QD theory for the fixed-point limit of \(r\) (assuming \(\omega\neq6\)): \[r_{FP}=1+\frac{\omega G_{1}(X_{FP})}{X_{FP}G'_{1}(X_{FP})(1-\nicefrac{\omega}{6})}. \label{rfp}\] Note that this expression is valid for Cases 1, 2 and 4 (not \(X=0\)), provided \(\omega\neq6\) and \(G'_{1}(X)\neq0\). Unlike QD theory, since \(\alpha_{\sigma}\neq 0\) we can use to obtain , which gives \(r\) in terms of the dynamical variables. If the system is to evolve towards its expected fixed point, \(r\) must approach \(r_{FP}\). Therefore, at the fixed point \(z^2\equiv\left(\frac{y}{G_{2}(X)}\right)^2\) should approach \[\begin{aligned}
\label{yg2} z_{FP}^2&\equiv&\left(\frac{y_{FP}}{G_{2}(X_{FP})}\right)^2\\ &=&1-\alpha_{\sigma}\frac{G_{1}(X_{FP})(1-\nicefrac{\omega}{6})(G'_{1}(X_{FP}))^{2}}{\left[X_{FP}G'_{1}(X_{FP})(1-\nicefrac{\omega}{6})+\omega G_{1}(X_{FP})\right]^2}\.\nonumber \end{aligned}\] In order for this approach to be smooth, one also needs \[\frac{dz}{dN}=-\left[4+\frac{G_2'(X)}{G_{2}(X)}\frac{dX}{dN}\right]z\to0.\] One possibility is that \(z\to0\), which requires a fine tuned relation among the parameters of the model. Since this is subsumed in Case 2 anyway, we will not analyze this particular case any further. A second possibility is that the quantity in square brackets approaches zero in the fixed-point limit. This implies a constraint equation for the dynamical variables near the fixed point. We note that in previous literature this constraint equation has been assumed to hold for the full dynamics with no justification [^6]. In such a case, the evolution of the dynamical system near the fixed point could be described in terms of 4 instead of 5 dynamical variables; i.e., near the fixed-point limit the evolution would be driven by the same dynamical equations as in the QD theory. However, *a priori* there is no reason to expect this condition to be valid for \(\alpha_{\sigma}\neq0\).
## Case 2: \(\left[\frac{\dot{\phi}^0}{n}\right]_{FP} =0\)
For Case 2, one has to solve to get the fixed-point value of \(X\). We must take care that, after solving for \(X_{FP}\), the sign of \(6(\tilde{h}^2_{FP}-G_{1}(X_{FP}))+X_{FP}G'_{1}(X_{FP})\) should be the same as the sign of \(G'_{1}(X_{FP})\)[^7]. It is worthwhile noting that for Case 2, \(\alpha_\sigma>0\). (We omit the QD case, \(\alpha_\sigma=0.\)) This can be seen by inspection of equation , recalling that \[\tilde{\Omega}_{DE}-G_1(X)=\frac{\omega}{6}\left(\frac{\dot\sigma}{m_gM_{\rm Pl}}\right)^2\geq0\. \nonumber\]
## Case 3: \(X_{FP}=0\)
From and the definition of \(\tilde{\Omega}_{DE}\), and , we get \[\label{sigdotx0} \dot{\sigma}_{FP}=0\.\] We find that \(r_{FP}\) is indeterminate from both and . Examining , we see that there are two ways to get \(X_{FP}=0\). The first possibility is that \(\sigma\to-\infty\), in which case we can draw no conclusion from about the value of \(\frac{\dot{\phi}^{0}}{n}\) at the fixed point. The second possibility is that \(n^{2}=(\dot{\phi}^{0})^{2}\) at the fixed point.
## Case 4: \(X_{FP}=1\)
Since \(G_{1}(1)=0\) and \(G_{1}'(1)=1\), and since \(\Omega_{m}=\Omega_{r}=0\) at the fixed point, from and , we find that \(X_{FP}=1\) requires \(\omega=6\). Substituting \(X=X_{FP}=1\) in , we get \[r_{FP}=1+6\tilde{h}_{FP}^{2} \label{rx1}\] where \(\tilde{h}_{FP}\) is indeterminate. Therefore the theory loses its predictive power. For this reason we will not consider this case anymore in the following analysis.
# Fixed-Point Linear Stability {#stability}
We wish to check whether or not the fixed-point solutions are attractors in the asymptotic future. This would be the case if any small perturbation around the fixed point decays to zero asymptotically. We start with the prescription given by to evaluate the fixed-point stability. Let \(\mathbf{V}=\left[{y},{X},{\tilde{\Omega}_{DE}},{\tilde{\Omega}_m},{\tilde{\Omega}_r}\right]^{T}\) denote the dynamical variables and \(\mathbf{f}(\mathbf{V})\) be the RHS of the first order differential equations. Thus we can express equations - as \[\frac{d\mathbf{V}}{dN}=\mathbf{f}(\mathbf{V})\,.\] Assuming small perturbations \(\delta{\mathbf{V}}\) around any point, \(\mathbf{V}_{0}\), a Taylor expansion of the functions \(\mathbf{f}(\mathbf{V})\) gives us \[\label{taylor} \frac{d}{dN} \mathbf{\delta V}={\rm \bf M} \mathbf{\delta V} + \mathbf{f}\left(\mathbf{V}_{0}\right)\,,\] where \({\bf M}\) is the stability matrix. Its elements are given by \[\begin{aligned}
\label{stabmat} M_{ij}&=&\left[\frac{\partial f_{i}(\mathbf{V})}{\partial V_{j}}\right]_{\mathbf{V}=\mathbf{V_0}}, i,j=1,\dots,5 \end{aligned}\] In Appendix [\[StandardStability\]](#StandardStability){reference-type="ref" reference="StandardStability"}, we provide the analytical expressions for the elements of \({\rm \bf M}\) (Equations -). If \(\mathbf{V}_{0}\) is a fixed point, then the second term of RHS of would vanish. Using the eigenvectors of \({\rm \bf M}\), one can then find matrix \({\rm \bf P}\) such that \[\begin{aligned}
{\rm\bf\mathcal{D}}&=&{\rm\bf P}^{-1}{\rm\bf M}{\rm\bf P},\label{diagonal}\\ &=&\text{Diag}(\lambda_{1},\dots,\lambda_{5}) \end{aligned}\] where \((\lambda_{1},\dots,\lambda_{5})\) are the eigenvalues given by . We define \(\delta \mathbf{\bar{V}}\) as \[\begin{aligned}
&{\rm\bf P} \delta \mathbf{\bar{V}}\equiv\delta\mathbf{{V}},\label{PdelV} \end{aligned}\] Thus the solution to in the new basis would be \[\delta{\bar{V}_{i}}=e^{\lambda_{i}N}C_{i}.\] where \(C_{i}\)'s are integration constants. Multiplying the above equation by \({\bf P}\) and thereby returning to the original basis, we get \[\delta{V}_{i}=\sum_{j=1}^{5}P_{ij}e^{\lambda_{j}N}C_{j}. \label{pert_soln}\] Then, for the fixed points to be stable, we require \(\delta{\mathbf{V}}\) to approach zero as \(N\to\infty\). It can be seen from that if the eigenvalues of \({\rm\bf M}\) are either real and negative or imaginary with negative real part, the fixed points will be stable or form a stable spiral respectively. We find in Appendix [\[StandardStability\]](#StandardStability){reference-type="ref" reference="StandardStability"}, that to obtain attractor solutions, \(\lambda_{4}\) and \(\lambda_{5}\) must be real and negative or complex with negative real parts. This requires the elements of \({\rm\bf M}\) to satisfy the condition \[\label{stable} 0\leq(3+2M_{22})^2+4M_{23}M_{32}<9\] for stable solutions or \[\label{stablespiral} (3+2M_{22})^2+4M_{23}M_{32}<0\] for stable spiral solutions.
## Case 1: \(J(X_{FP})=0\) {#case1 stability}
The above-described standard method will not suffice to evaluate fixed-point stability in all cases. In particular, for \(J(X_{FP})=0\) the stability matrix \({\rm\bf M}\) has divergent terms at the fixed point (see Appendix [\[StandardStability\]](#StandardStability){reference-type="ref" reference="StandardStability"} for the exact expressions of its elements). This makes the evaluation of the matrix \({\rm\bf P}\) and its inverse indeterminate. One can thus no longer diagonalize the system of equations as in and come up with solutions for given by . We devise the following scheme to assess stability: we introduce small perturbations \(\delta{\mathbf{V}}\) around an arbitrary point \(\mathbf{V}_{0}\); using and diagonalizing \({\rm\bf M}\), we solve for \(\delta{\mathbf{V}}\). In order for the fixed point to be an attractor we require the following conditions to be satisfied. *If \(\mathbf{V}_{0}\) is infinitesimally close to the fixed point \(\mathbf{V}_{FP}\), then, as \(N\to\infty\)*:
- *the perturbations \(\delta{\mathbf{V}}\) are infinitesimally close to zero,* therefore we require \[\begin{aligned}
\lim_{\mathclap{\substack{N\to+\infty \\ V_{0}\to V_{FP}}}}\quad \mathbf{\delta V} \to 0 \label{deltaV}\,; \end{aligned}\]
- *the derivatives of perturbations, \(d\, \delta{\mathbf{V}}/dN\), are infinitesimally close to zero,* therefore we must verify that \[\begin{aligned}
\lim_{\mathclap{\substack{N\to+\infty \\ V_{0}\to V_{FP}}}}\quad \frac{d}{dN} \mathbf{\delta V} \to 0 \label{dNdeltaV}\,; \end{aligned}\]
- *\(y/G_{2}(X)\) is infinitesimally close to its fixed-point limit --* in compact form this translates to \[\begin{aligned}
\label{ybyg2} \lim_{\mathclap{\substack{N\to+\infty \\ V_{0}\to V_{FP}}}}\quad \left(\frac{y}{G_{2}(X)}\right)^{2} \to \qquad\qquad\qquad\qquad\qquad \, \\ 1-\alpha_{\sigma}\frac{G_{1}(X_{FP})(1-\nicefrac{\omega}{6})(G'_{1}(X_{FP}))^{2}}{\left[X_{FP}G'_{1}(X_{FP})(1-\nicefrac{\omega}{6})+\omega G_{1}(X_{FP})\right]^2}\,. \nonumber \end{aligned}\]
In more detail, after diagonalizing the matrix **M**, from we obtain \[\begin{aligned}
\frac{d}{dN} \mathbf{\delta \bar{V}}=\mathcal{D}\mathbf{\delta \bar{V}}+ \mathbf{B}, \label{taylor2} \end{aligned}\] where \[\begin{aligned}
&\mathbf{B}=P^{-1}\mathbf{f}\left(\mathbf{V}_{0} \right),\label{B}\\ \nonumber \end{aligned}\] Solving , we are left with \[\begin{aligned}
\mathbf{\delta \bar{V}}&=\text{Diag}\left(-\frac{1}{\lambda}_{1},\cdots,-\frac{1}{\lambda}_{5}\right)\mathbf{B}\\ \nonumber &+\text{Diag}\left(e^{\lambda_{1} N},\cdots,e^{\lambda_{5} N}\right){\mathbf{C}}\,, \end{aligned}\] where \(\mathbf{C}\) is the integration constant vector. Upon returning to the original basis, we find \[\begin{aligned}
\label{perturbed} \mathbf{\delta {V}}&={\rm\bf P}\hspace{1mm} \text{Diag}\left(-\frac{1}{\lambda}_{1},\cdots,-\frac{1}{\lambda}_{5}\right){\rm\bf P}^{-1}\mathbf{f}\left(\mathbf{V}_{0}\right)\\ \nonumber &+{\rm\bf P}\hspace{1mm}\text{Diag}\left(e^{\lambda_{1} N},\cdots,e^{\lambda_{5} N}\right){\mathbf{C}}. \end{aligned}\] In Appendix [\[ModifiedStability\]](#ModifiedStability){reference-type="ref" reference="ModifiedStability"} we show that the requirement (A) could be satisfied if the eigenvalues \(\lambda_{i}\)'s are either real and negative or complex with negative real part. It is also shown that, if \(X_{0}\) is a point infinitesimally close to its fixed-point value \(X_{FP}\), it must satisfy the following relations[^8]
1. \(\alpha_{\sigma}>0\) \[\begin{aligned}
G'_{1}(X_{0})G_{2}(X_{0})G'_{2}(X_{0})&>&0\label{reln1},\hskip 2pt 0<\omega<6 \\ G'_{1}(X_{0})G_{2}(X_{0})G'_{2}(X_{0})&<&0\label{reln2},\hskip 2pt \omega>6 \,, \end{aligned}\]
2. \(\alpha_{\sigma}<0\) \[\begin{aligned}
G'_{1}(X_{0})G_{2}(X_{0})G'_{2}(X_{0})&<&0\label{reln3},\hskip 2pt 0<\omega<6 \\ G'_{1}(X_{0})G_{2}(X_{0})G'_{2}(X_{0})&>&0\label{reln4},\hskip 2pt \omega>6\,. \end{aligned}\]
In the numerical investigation performed in Section [7](#scanning){reference-type="ref" reference="scanning"}, we will take \(X_{0}\) as the initial condition for the dynamical variable \(X\). Therefore, relations - provide the viable initial conditions needed to have linear stable solutions for the dynamical variables. We note here that - can only be satisfied for \(X_{0}\) either greater or less than \(X_{FP}\) and never both[^9]. For \(\mathbf{V_{FP}}\) to be an attractor, we are required to verify under which conditions (B) and (C) are satisfied. However, as explained in more details in Appendix [\[ModifiedStability\]](#ModifiedStability){reference-type="ref" reference="ModifiedStability"}, in linear perturbation theory, (B) and (C) cannot be determined. Therefore a numerical investigation on both and should be performed. In practice the analysis of \(\left(\frac{y}{G_{2}(X)}\right)^{2}\) is enough. Indeed we find that, even though the dynamical variables approach their fixed-point values in the asymptotic future, \(\left(\frac{y}{G_{2}(X)}\right)^{2}\) oscillates with maxima that *grow* in time. Eventually, it reaches the critical value of 1, making \(r\) in eq. singular. We discuss this phenomena in greater detail in Section [7](#scanning){reference-type="ref" reference="scanning"}.
## Case 2: \(\left[\frac{\dot{\phi}^0}{n}\right]_{FP} =0\)
In this Case, all elements of the stability matrix are well defined and finite in the fixed-point limit. Hence we use the standard approach of fixed-point analysis. The analytical form of the elements of \({\rm \bf M}\) around the fixed-point solutions are provided in Appendix [\[StandardStability\]](#StandardStability){reference-type="ref" reference="StandardStability"} (-). For the fixed points to be attractors, the elements of \({\rm\bf M}\) must satisfy the conditions for stable solutions and for stable spiral solutions.
## Case 3: \(X_{FP}=0\)
Recalling , we find that the presence of terms \(\frac{X}{\tilde{\Omega}_{\sigma}}\) and \(\frac{y}{G_{2}(X)}\) in the stability matrix make the elements , , , and and consequently the eigenvalues indeterminate. Hence the stability of \(X_{FP}=0\) is unclear.
# Numerical Investigation of fixed-point stability {#scanning}
As noted above, fixed-point linear stability conditions are necessary but not sufficient to guarantee that a fixed point can be reached by evolving from an initial configuration that is displaced from that fixed point. Moreover, as described above, for Case 1 and Case 3 some or all the relevant quantities that appear in the background equations cannot be analytically assessed in linear perturbation theory. Therefore we need to perform numerical tests of the fixed-points stability. For suitable values of the parameters \((\alpha_{3},\alpha_{4},\omega,\alpha_{\sigma})\), we check numerically that the 5 dynamical variables approach their fixed-point values if initially perturbed from them. Crucially, this includes verifying that \(z=y/G_{2}(X)\) approaches the value given by . In our numerical investigations, we use the results of Section to guarantee linear stability. We look further for values of the parameters for which:
(i) \(X_{FP}\geq0\); [\[check1\]]{#check1 label="check1"}
(ii) \(\tilde{h}_{FP}^2\) given by is positive; [\[check2\]]{#check2 label="check2"}
(iii) the fixed-point linear stability conditions (given by - for Case 1, and or for Case 2) are satisfied. [\[check3\]]{#check3 label="check3"}
Note that we could further constrain the parameter space by selecting regions where the scalar, vector and tensor perturbations are stable . Although it is not necessary for the present analysis, to simplify our search, we use some of the necessary restrictions imposed by the stability of the perturbations, which we summarize in Appendix [\[perturbations\]](#perturbations){reference-type="ref" reference="perturbations"}. As discussed below, we find that the traditional (Case 1) fixed points cannot be reached from any neighboring configuration because \(z\) does not approach the corresponding fixed-point value! The Case 2 fixed points behaved as expected from the linear stability analysis. The Case 3 fixed points always encounter a singularity before reaching the fixed-point values.
## Case 1: \(J(X_{FP})=0\) {#numCase1}
Our goal in this section is nominally to identify values of the parameters \(({\alpha_{3},\alpha_{4},\omega,\alpha_{\sigma}})\) such that - hold true and verify that the dynamical variables approach their fixed-point values in the asymptotic future. One of the central results of this paper is that we fail to do so. We show that, for all the choices of parameters and initial conditions we consider, the dynamical variables *never* approach their fixed-point values if they do not start there to begin with. We begin the discussion with an example by setting \(({\alpha_{3},\alpha_{4},\omega,\alpha_{\sigma}})=(3.1, 3.1, 5.5, 10.0)\). Solving , we find that \(X_{+}\) satisfies the conditions -. This set of parameter values also satisfies the tensor, vector and scalar stability at the \(X_{+}\) fixed point (see Appendix [\[perturbations\]](#perturbations){reference-type="ref" reference="perturbations"}). We remind the reader that since \(G_{2}(X)\) vanishes at the fixed point, the stability conditions - only hold true in the vicinity of the fixed point. In this example, since \(0<\omega<6\) and \(\alpha_{\sigma}>0\), the initial value of \(X\) must be of the form \(X_{0}=X_{+}+\epsilon\), where \(0<\epsilon\ll1\) so that is satisfied. To study the behavior of the dynamical variables close to the fixed point, we set \(\epsilon={10^{-6}}\) and the initial conditions to be \({\tilde{\Omega}}_{DE, 0} = {\tilde{\Omega}}_{DE, FP}+\epsilon\), \({\tilde{\Omega}}_{m, 0}=\epsilon\), \({\tilde{\Omega}}_{r, 0}=\epsilon\). We recall that for Case 1, we cannot freely perturb \(y\), because the initial value of \(r\) from would not then be close to its fixed-point value . Instead, we perturb \(r\) by \(\epsilon\) and initialize \(y\) using . For a detailed discussion, we refer the reader to Appendix [\[ModifiedStability\]](#ModifiedStability){reference-type="ref" reference="ModifiedStability"}. Notice that \({\tilde{\Omega}}_{DE, 0}\), \(r_{0}\), \(y_{0}\) could be either greater or less than the fixed-point values. In this example we chose the former. After setting the initial conditions as described above, we study the behavior of the dynamical variables with time. Equations , and have simple solutions, \[\begin{aligned}
\label{yOmOr} y=y_{0}e^{-4N},\hskip 2pt\tilde{\Omega}_{m}=\tilde{\Omega}_{m,0}e^{-3N},\hskip 2pt\tilde{\Omega}_{r}=\tilde{\Omega}_{r,0}e^{-4N}\. \end{aligned}\] The evolution equations for \(X\) and \(\tilde{\Omega}_{DE}\) have no analytical solutions. Therefore, we use in and , and solve equations and numerically. We find that and evolve until they reach a singularity. As shown in Fig. [\[dyn_var\]](#dyn_var){reference-type="ref" reference="dyn_var"}, the evolution of \(X\) and \(\tilde{\Omega}_{DE}\) before the singularity turns out to be exactly what we expect from perturbation theory: they oscillate around their fixed-point values with decaying amplitude and increasing frequency. By inspecting the RHSs of and we can identify the possible sources of the singular behavior. The only possibilities are that, as the fixed point is approached, the square-root term \(\left(\tilde{\Omega}_{DE}-G_{1}(X)\right)\) in or \(\left(1-\left(\frac{y}{G_{2}(X)}\right)^{2}\right)\) in approaches zero. From the *top panel* of Fig. [\[zsq_full\]](#zsq_full){reference-type="ref" reference="zsq_full"} we see that \(\left(\tilde{\Omega}_{DE}-G_{1}(X)\right)\) does not vanish as the fixed point is approached. On the other hand in Fig. [\[zsq_both\]](#zsq_both){reference-type="ref" reference="zsq_both"} we show that \(z^{2}=\left(\frac{y}{G_{2}(X)}\right)^{2}\) exhibits an oscillating behavior where the peaks grow in time monotonically. Thus \(z\) does not approach its expected fixed-point value. Moreover, as shown in the *bottom panel* of Fig. [\[zsq_full\]](#zsq_full){reference-type="ref" reference="zsq_full"}, it eventually approaches unity, causing a breakdown in the coupled dynamical equations for \(X\) and \(\tilde{\Omega}_{DE}\). To ensure that the phenomenon described above is not a numerical artifact, we performed numerous numerical and analytic tests. Numerically, we confirmed that the \(z^2\to1\) behavior was robust to: increasing the numerical precision demanded from the integrator, increasing the order of the integrator, changing the integration scheme, and changing the initial conditions. Therefore, we have conclusive evidence that the evolution of the dynamical variables drives \(z^{2}\) towards unity. We therefore conclude that for this value of the EQDMG parameters, in the neighborhood of a Case 1 fixed point the evolution equations and reach a singularity in the asymptotic future. In order to confirm that this conclusion holds true for all regions of the parameter space we would need to evolve the differential equations for a high number of different values of \(({\alpha_{3},\alpha_{4},\omega,\alpha_{\sigma}})\) and initial conditions close to the fixed-point solutions. This is numerically too costly, therefore we solve the exact equations for a limited number of points in the parameter space (few thousands). For a more extensive scan (millions of points) we rely on a sensible approximation based on perturbation theory and on the following argument. If the fixed point is an attractor, perturbation theory should be increasingly accurate the closer the initial conditions are to the fixed-point value; if the fixed point is not an attractor, then perturbation theory may or may not work. If in perturbation theory we were to find that \(z\) approached its fixed-point value, then we would want to verify that conclusion with the full non-perturbative solution. If \(z\) does not approach its fixed-point value in perturbation theory, then that may be because the initial perturbation away from the fixed point was too large --outside the basin of attraction--so one should decrease the magnitude of the perturbation as much as possible. The smaller the perturbation, the more one would expect to trust perturbation theory. If this still fails, it is always possible that the basin of attraction is extremely small and difficult to find numerically; but in any case, while we will not have arrived at a mathematical proof that this fixed point is not an attractor, we will certainly have shown that it is not a suitable candidate for a cosmological model. We start by testing the prediction of \(z^{2}\) from perturbation theory in the previously considered example. We rely on equations , , , , where the \(C_{i}\) coefficients are determined by requiring that \(\mathbf{\delta {V}}(N_{in}\equiv0)=0\). In Fig. [\[zsq_both\]](#zsq_both){reference-type="ref" reference="zsq_both"} we plot the behavior of \(z^{2}\) for the first \(\sim{20}\) periods comparing the full and perturbation theory solutions. We verify a consistent growth in the peaks of \(z^{2}\). Hence, in the subsequent analysis, we employ the following supporting argument: if in perturbation theory the first \(20\) peaks of \(z^{2}\) grow monotonically with time, we conclude that \(z^{2}\) does not approach its fixed-point value. We randomly selected \(\sim{10^7}\) values of the parameters in the following ranges \[\begin{aligned}
\label{region2} 10^{-5}&<\left|\alpha_{3}\right|,\left|\alpha_{4}\right|,\alpha_{\sigma}<10^{5},\nonumber \\ 0&<\omega<6 \text{ or } 6<\omega<10^{5} \end{aligned}\] and evaluate \(X_{\pm}\). We evaluate the first \(\sim{20}\) peaks of \(z^2\) for all points that satisfied the conditions - together with the pertinent ones in Appendix [\[perturbations\]](#perturbations){reference-type="ref" reference="perturbations"} using linear perturbation theory. We set the initial conditions of the dynamical variables to be from \(\epsilon={10^{-8}}\) to \(\epsilon={10^{-12}}\) away from their fixed-point values[^10]. We repeated the same procedure using the numerical solutions to the full (non-perturbative) equations with few thousand randomly selected points. We find that the peaks of \(z^2\) always grow monotonically in both linear perturbation theory and using the exact equations. Notice that for \(\alpha_{\sigma}<0\), equations and imply \(z^{2}>1\). However, even for negative values of \(\alpha_{\sigma}\), \(r\) from still becomes singular because the troughs of \(z^{2}\) monotonically decrease and approach unity. Hence we come to the following conclusion: *the standard (Case 1) fixed-point solutions of Extended Quasi-Dilaton Massive Gravity are dynamically unattainable due to an unavoidable singularity while approaching them. Hence, we rule out the suitability of this fixed point as a cosmological model.*
## Case 2: \(\left[\frac{\dot{\phi}^0}{n}\right]_{FP} =0\)
Recalling that \(\alpha_{\sigma}\) is positive, we randomly select points in the ranges . To obtain the fixed-point values of \(X\), we use and to arrive at a 7th order polynomial in \(X\) with no analytical solutions. After numerically evaluating the roots of the polynomial, we demand that the fixed-point values satisfy - and that the vector and tensor perturbations are stable[^11](see Appendix [\[perturbations\]](#perturbations){reference-type="ref" reference="perturbations"} for details). We find parameter values for which both stable and spiral solutions are allowed. We selected a few points in the allowed parameter space to verify the attractor behavior of the fixed points. Starting from small perturbations (\(\epsilon =10^{-6}\)) around the fixed-point values of the dynamical variables, we study the evolution of the differential equations with time by solving , and analytically and and numerically. As predicted from linear stability analysis (Section [6](#stability){reference-type="ref" reference="stability"}), we find that the dynamical variables reach their fixed-point values in the asymptotic future.
## Case 3: \(X_{FP}=0\)
For the Case 3 fixed points we do not have indications from the linear stability analysis. We perform a random search selecting \(64\times10^{4}\) points in the parameter space. Starting from small perturbations (\(\epsilon =10^{-6}\)) around the fixed-point values we find that the dynamical variables are approaching the fixed-point limit. However, we find that the quantity \(\left(\tilde{\Omega}_{DE}-G_{1}(X)\right)\) always approaches zero at finite time. We find numerically that this makes \(r\) vanish which is not allowed in the theory.
# Perturbations: Avoiding the Boulware-Deser Ghost {#bdghost}
One of the outstanding concerns in any theory that adds a scalar degree of freedom to Einstein-Hilbert gravity is the possibility that the theory includes a Boulware-Deser ghost-- a dynamical degree of freedom with a wrong-sign kinetic term. Due to disagreements in the literature mentioned above (see ), in this section we reconsider the analysis of the scalar perturbations, and determine under what conditions there is or not a necessary additional constraint equation that generically allows one to eliminate the Boulware-Deser mode. We will show that such a constraint exists when \(J(X)=0\) (Case 1) or \(\dot{\sigma}=0\), giving the possibility of being ghost-free at the corresponding fixed points. In both cases, there is also a well-defined set of initial conditions of the dynamical variables for which this virtue extends beyond the fixed point. The first case is the only one potentially relevant for cosmology, and we explore the consequences of the initial conditions below. The second case \(\dot{\sigma}=0\) corresponds to a Minkowski background metric (provided \(X\neq 0\)) which we consider only in Appendix [\[Minkowski\]](#Minkowski){reference-type="ref" reference="Minkowski"}. Given that in the previous section, for Case 3 we have shown \(X=0\) cannot be approached dynamically from any neighborhood, one might consider the possibility when \(X(N)=0\) \(\forall\) \(\rm{N}>0\). This possibility can also be ruled out because \(X=0\) only makes sense as a fixed-point limit. Hence from now on, we will not consider Case 3 any further. Following the standard treatment and in this section only, we take the action of the matter sector to be [^12] \[\label{sm} S_m= \int d^4x \sqrt{-g} \,P(Y,\chi)\,,\] which corresponds to the addition of a scalar field \(\chi\) with a non-canonical kinetic term, given by the function \(P(Y,\chi)\), where \[Y \equiv-\frac{1}{2}\partial_\mu\chi\partial^\mu\chi\.\] The fluid variables (pressure \(p\), energy density \(\rho\) and sound speed \(c_s\)) associated with \(\chi\) can be written as: \[p = P (Y,\chi) \,,\qquad \rho \equiv 2P_{,Y} Y-P, \qquad c_s^2 \equiv \frac{P_{,Y}}{\rho_{,Y}}\. \label{eq:fluid}\] To study the perturbations in \(\chi\), we replace \(\chi\) by \(\chi+M_{\rm Pl}\delta\chi\). To facilitate comparisons, we adopt notations as close as possible to those most used in the literature. We decompose the metric into tensor, vector, and scalar as \[\begin{aligned}
\delta g_{00} &=&-2\, \Phi\,,\\ \delta g_{0i} &=& \,a\,\left(\partial_i B+B_i\right)\,,\nonumber\\ \delta g_{ij} &=& a^2 \left[2\,\delta_{ij}\psi +\left(\partial_i\partial_j-\frac{\delta_{ij}}{3}\partial^k\partial_k\right)E+\partial_{(i}E_{j)}+h_{ij}\right]\,,\nonumber \end{aligned}\] where \(\delta^{ij}h_{ij} = \partial^ih_{ij} = \partial^i E_i = \partial^i B_i=0\) and Latin indices are raised with \(\delta^{ij}\). For the quasi-dilaton field we write the background plus perturbation by replacing \(\sigma\) by \(\sigma + M_{\rm Pl}\delta\sigma\). For the sake of clarity, in this section we present only those results that are either more generic or different from those in previous literature, and relegate the remaining details to Appendix [\[AppendPerturb\]](#AppendPerturb){reference-type="ref" reference="AppendPerturb"}. We focus first in the case \(\dot{\phi^0}\neq 0\) and, following literature, we adopt the unitary gauge \(\phi^{a}=\phi^{0}(t)\delta^a_0 +\delta^a_ix^i\). The case \(\dot{\phi^0}(t)= 0\), for which we cannot work in this gauge, is analyzed later.
## \(\dot{\phi^0}\neq 0\)
The part of the action ([\[eq:action\]](#eq:action){reference-type="ref" reference="eq:action"}) that is quadratic in the perturbations can be split into tensor, vector and scalar contributions. We focus here on the scalar part. From the variation of the quadratic action for the scalar sector with respect to \(\Phi\) and \(B\), we obtain constraint equations that allow us to eliminate these two variables. The solution for \(\Phi\) and \(B\) can be found for a generic background and without making use of the background equations[^13]. Introducing the solutions for \(\Phi\) and \(B\) back into the action, we can write the kinetic part as \[\label{kineticAction} S^{(2)}_{\rm scalar}\supset \frac{M_{\rm Pl}^2}{2}\int d^3k \,dt\,a^3 \dot{Z}^\dagger\,\mathcal{Q}\,\dot{Z}\,,\] with \(Z=\{\psi,\delta\sigma, E,\delta\chi\}.\) One combination corresponds to the Boulware-Deser mode. For this mode to be non-dynamical, the determinant of the matrix \(\mathcal{Q}\) must vanish. We first consider the vacuum case, and then see how the inclusion of matter changes the results.
- Vacuum case: In the absence of the additional field \(\chi\), \(\mathcal{Q}\) is a \(3\times3\) matrix and \(Z=\{\psi,\delta\sigma, E\}.\) The determinant \({\rm Det}(\mathcal{Q})\) can be computed analytically in Fourier space. After expanding in powers of comoving wavenumber \(k\), keeping only the leading order terms (the infrared part), and using the background equations to express \({H}\) and \(\dot{H}\) in terms of the other dynamical variables, \[\begin{aligned}
& {\rm Det}(\mathcal{Q})=\\ \nonumber &\frac{4 \omega \alpha_\sigma M_{\text{Pl}}^4 a^2 J(X) {\dot{\phi^0}}^2 \dot{\sigma}^2 k^4} {r^3 X m_g^2 \left(J(X)+\alpha_3 (X-1)^2-3 X+3\right)} +\mathcal{O}(k^6)\.\nonumber \end{aligned}\] The fact that the background equations are satisfied does not guarantee that the determinant vanishes, as previous computations suggested. However, it is clear that \({\rm Det}(\mathcal{Q})\) vanishes when \(J(X)=0\) or \({\dot{\sigma}}=0\). Moreover, in that cases it can be shown that the determinant vanishes at all order in \(k\). As we show next, these results are robust to the addition of matter.
- With Matter: Proceeding as above, at leading order in the wavenumber \(k\), after using the background equations to replace \(\dot{H}\) and \({H}\) in terms of the other dynamical variables, the determinant of the now \(4\times 4\) matrix \(\mathcal{Q}\) can be written as \[\begin{aligned}
&&{\rm Det}(\mathcal{Q})= \frac{32 \omega }{ r^3 X } \alpha_{\sigma } M_{\text{Pl}}^6 a^2 J(X)\dot{\phi^0}^2 (X-1) P_{,Y}(Y,\chi) {\dot{\sigma}}^2 k^4 \nonumber\\ &&\times \Big{\{}p+\rho(1-2 c_s^2)-2 c_s^2 (X-1) m_g^2 M_{\text{Pl}}^2 \nonumber\\ &&\times \left(J(X)+\alpha _3 (X-1)^2-3 X+3\right) \Big{\}}^{-1} +\mathcal{O}(k^6)\. \end{aligned}\] Therefore, ignoring the particular cases for which \(X=1\), we see that the determinant vanishes in the infrared under the same conditions as in the vacuum case. Under those conditions, we also check the determinant vanishes at all order in \(k\). Of course the determinant also vanishes in the case where the matter is just a cosmological constant, \(P_{,Y}=0\).
Therefore, \({\rm Det}(\mathcal{Q})=0\) at the Case 1 fixed points. If we perturb \(X\) away from that fixed-point value, then \(J(X)\neq0\), and \({\rm Det}(\mathcal{Q})\neq0\) either. However, from , we can conclude that if \(J\equiv J(X)=0\) at some initial time \(t_0\), then also \(\dot{J}=0\) at that time, and also \(\ddot{J}\). Thus \(J(X)\) remains zero at all times once \(X=X_{FP}\). Consequently \({\rm Det}(\mathcal{Q})\) also remains equal to zero. Therefore, in order to eliminate the BD ghost in both the vacuum and matter contexts, we must impose special initial conditions on the dynamical variables; namely, that \(X=X_\pm\) is exactly satisfied. Setting \(X\) to this fixed-point value, and thus setting \(J(X)=0\), \(y=0\) and \(\dot{X}=0\), we can solve through to obtain: \[\begin{aligned}
\label{jxcond} \tilde{\Omega}_{DE}\ &=& \frac{\frac{\omega}{6}\left(\tilde{\Omega}_{m0}e^{-3N}+\tilde{\Omega}_{r0}e^{-4N}\right)+ G_{1}(X_\pm)}{1-\frac{\omega}{6}}\nonumber\\ \tilde{h}^2 &=& \frac{6}{\omega}\left(\tilde{\Omega}_{DE}-G_1(X_\pm)\right) \end{aligned}\] This appears to be just a fine-tuning of the dynamical variable \(X\) to some parameter-dependent value; however \(X\) is a function of both the quasi-dilaton and the scale factor given by . This thus appears to be an awkward fine-tuning, relating the initial values of many of the dynamical variables to one another. It also allows the fixed points to become stable attractors that can be approached dynamically. Furthermore, notice that in the asymptotic past, \(N\to-\infty\), the matter and radiation terms will dominate over \(G_{1}(X\pm)\). This restricts the values of \(\omega\) to \(0<\omega<6\) or it means that the theory is not valid arbitrarily far into the past.
## \(\dot{\phi^0}= 0\)
As mentioned earlier, in the special case \(\dot{\phi^0}= 0\) (Case 2) we cannot use the unitary gauge. Assuming \(H\neq 0\) we choose the gauge with \(\psi=0\) instead, while we keep \(\phi^{i}=x^i\). In principle, as first noticed by , this case could also be interesting beyond the fixed-point limit. This corresponds to setting \(\dot{\phi^0}(t_0)= 0\) at some initial time \(t_0\), yielding \(\ddot{\phi^0}(t_0)= 0\) and \(\dot{\phi^0}(t)= 0\) for all time. This enforces \(y=0\), but the other dynamical variables remain free to evolve. The perturbations of the Stückelberg scalar degree of freedom \(\delta\phi^0\) (since \(\dot{\phi^0}=0\) at the background level) only enter as a contribution to \(f_{\mu\nu}\) that is quadratic in \(\delta\phi^0\). Therefore, the quadratic action for this scalar decouples from the other parts, and the kinetic part can be immediately computed \[S^{(2)}_{\rm scalar} \supset-\frac{M_{Pl}^2}{2} \int d^3k \,dt\,a^3 \,\frac{m_g^2}{r} G_2(X) a^2|\delta\dot{\phi^0}|^2\.\] Now, to compute the determinant of the kinetic matrix corresponding to the other degrees of freedom we proceed as above. We integrate out \(\Phi\) and \(B\), we write the relevant part of the action as in ([\[kineticAction\]](#kineticAction){reference-type="ref" reference="kineticAction"}), and we consider the vacuum case and the case with matter separately:
- Vacuum case: In this case \(Z=\{\delta\sigma,E\}\), and \[\begin{aligned}
&&{\rm{Det}}(\mathcal{Q})=G_1'(X) k^4 m_g^2 X \omega a^2 \left(6 \alpha_{\sigma} G_1(X)+r^2 X^2 \omega \right)\nonumber\\ &&\times\Big{\{} 4 \left(9 \alpha_{\sigma}G_1(X) G_1'(X) m_g^2 X a^2+k^2 (r+1) \left(6 \alpha_{\sigma} G_1(X)\right.\right.\nonumber\\ && +r^2 X^2 \omega ))\Big{\}}^{-1}; \end{aligned}\]
- With matter: \(Z=\{\delta\sigma,E,\delta\chi\}\), and \[\begin{aligned}
&&{\rm{Det}}(\mathcal{Q})=-H^2 k^4 X \omega m_g^2 G_1'(X) M_{\text{Pl}}^2\nonumber \\ &&\times\left[X m_g^2 \left(r^2 X \omega-(r-1) \alpha _{\sigma } G_1'(X)\right)+2 \dot{H} \alpha _{\sigma }\right]\nonumber\\ &&\times \Big{\{}\dot{\chi }^2 \left[a^2 X m_g^2 G_1'(X) \left(2 \alpha _{\sigma } \left(3 H^2 c_s^2+\dot{H}\right)\right.\right.\nonumber\\ &&-\left.\left.r^2 X^2 \omega \left(c_s^2-1\right) m_g^2\right)-a^2 (r-1) X^2 \alpha _{\sigma } m_g^4 \left(G_1'(X)\right){}^2\right.\nonumber\\ & &+\left.4 H^2 k^2 (r+1) \alpha _{\sigma } c_s^2\right]\Big{\}}^{-1} \end{aligned}\]
Hence, we conclude that the condition \(\dot{\phi^0}=0\) is not sufficient to obtain the required additional constraint.
# Conclusions {#cnclsns}
In this paper we studied the fixed-point solutions of EQDMG in great detail after splitting them into four cases. We performed a linear stability analysis of the background (homogeneous) fixed-point solutions. This stability analysis for the standard case (Case 1) required an unconventional approach. We derived necessary stability conditions for the dynamical variables. However, we verify numerically that the dynamical variables inevitably encounter a singularity while approaching their fixed-point values. This is because a function of two of the dynamical variables fails to approach its fixed-point limit, and instead oscillates, finally reaching a critical value at which the dynamical equations are no longer valid. On the other hand, in Case 2, the dynamical variables smoothly asymptote towards their fixed-points values. However, in this case, the presence of the additional constraint that would allow one to eliminate the BD mode is not guaranteed. A numerical search of the Case 2 parameter space revealed many values of the parameters that have stable evolutions toward the fixed point. We analyzed the conditions under which the constraint equation that generically allows for the elimination of the Boulware-Deser ghost can be obtained. We found that the constraint equation exists in Case 1 type of solutions, but not around a generic background. Moreover, in these solutions the fixed points are attractors. However, such solutions require the time-derivative of the quasi-dilaton field must be exactly tuned against the Hubble parameter. The conclusive result of our study indicates that the EQDMG theory shows pathological behaviors when a generic FLRW solution is assumed. Only an "awkward\" fine-tuned solution is healthy. From our point of view this finding makes the EQDMG less appealing as a viable model to explain the evolution of our Universe. The extensive analytical and numerical analysis presented for the EQDMG theory must be carried out for all the proposed massive gravity theories that provide flat Friedmann-Lemaı̂tre-Robertson-Walker solutions (a non-exhaustive list includes ). As mentioned below , the condition \(\omega<6\) arises in EQDMG when the \(J(X)=0\) solution is taken to describe the past cosmological evolution, when matter and radiation dominated. In EQDMG this solution is imposed to avoid the presence of the BD ghost. However, since in QDMG the BD mode can be eliminated without setting \(J(X)=0\), the restriction \(\omega<6\) is in principle unnecessary. It would be worth exploring the parameter region \(\omega>6\) to see whether the simpler QDMG theory allows a proper description of the expansion history of the Universe. We propose to contribute to the understanding of this issue in the future. We thank especially C. de Rham for helpful insights and M. Fasiello, K. Hinterbichler, A. Tolley and S.-Y. Zhou for discussions. SA, SK and GDS are partially supported by CWRU grant DOE-SC0009946. The research leading to these results has received funding from the European Research Council under the European Community Seventh Framework Programme (FP7/2007-2013 Grant Agreement no. 279954) ERC-StG \"EDECS\". SA and GDS thank the CERN theory group for their hospitality. We acknowledge the use of the xAct-xPand package for Mathematica.
[^1]: Note that, using the constraint equation , one can show that the equation obtained by taking the variation of \(S_\sigma\) with respect to \(\sigma\) is not an independent equation.
[^2]: The special case \(H=0\), i.e. Minkowski space, is discussed in appendix [\[Minkowski\]](#Minkowski){reference-type="ref" reference="Minkowski"}.
[^3]: We will focus our attention below on the positive sign, because the negative sign leads to only an \(X=0\) fixed point.
[^4]: For \(\alpha_{\sigma}=0\), \(r\) cannot be determined from , because gets reduced to \(y=G_{2}(X)\). Thus, we can no longer use , and the above system of evolution equations is not well-equipped to handle this case. In fact, this limit gives us the Quasi-Dilaton theory and the evolution of the dynamical variables have been previously studied by and.
[^5]: A special case occurs when \(\alpha_{4}=\frac{2\alpha_{3}^{2}}{3}\), in which case \(G_{1}(X)=0=J(X)\). We are left with \[\label{g1g20} \left(X_{FP},r,_{FP},\omega\right) = \left(1+\frac{3}{{\alpha}_{3}},1+\frac{6{\tilde h}^2}{X_{FP}^{2}},6\right).\] This solution for \(\omega=6\) and \({\tilde h}^2\) is indeterminate because of simplifications occurring in . Therefore, on the \(\alpha_{4}=\frac{2\alpha_{3}^{2}}{3}\) hypersurface in the parameter space, the dynamical equations lose predictive power.
[^6]: This constraint is eq. (7) of.
[^7]: There is a special solution of where \(G_{1}(X)=G'_{1}(X)=0\). This is possible only when \(3+2\alpha_{3}+3\alpha_{3}^{2}-4\alpha_{4} = 0\) and the common root of \(G_{1}(X)\) and \(G'_{1}(X)\) is \(X_{FP}=(3+5\alpha_{3}+2\alpha_{4})/\left(2\alpha_{3}+2\alpha_{4}\right)\), which corresponds to \(\tilde{h}_{FP}=0.\) Thus we are returned to the special case \(h=0\) discussed in Appendix [\[Minkowski\]](#Minkowski){reference-type="ref" reference="Minkowski"}.
[^8]: In special cases where \(G_1(X_{FP})=0\), such as \(\omega=6\), or \(2\alpha_3^2=\alpha_4\), the eigenvalues involve ratios of zeros that we are unable to resolve, so we cannot determine the stability.
[^9]: We exclude particular cases for which in a neighborhood of \(X_{FP}\) not only \(G_{2}(X)\) changes sign but also either \(G'_{2}(X)\) or \(G'_{1}(X)\).
[^10]: For simplicity we only consider \({\tilde{\Omega}}_{DE, 0}\), \(r_{0}\), \(y_{0}\) larger than their fixed-point values. We numerically inspect few examples with initial conditions smaller than the fixed-point values and we find the same singular behavior.
[^11]: We keep this analysis agnostic to the stability conditions for scalar perturbations since, as we show below, the fact that \(\dot{\phi^0}=0\) does not guarantee the absence of the BD ghost.
[^12]: Clearly, the action (80) does not describe dust-like matter. However, we consider it only for simplicity because it will be enough for the purposes of this section.
[^13]: There are some disagreements in the literature regarding these results. In the appropriate limit, our results reduce to those of rather than those of. For more details see Appendix [\[AppendPerturb\]](#AppendPerturb){reference-type="ref" reference="AppendPerturb"}. | {'timestamp': '2017-06-07T02:09:01', 'yymm': '1706', 'arxiv_id': '1706.01872', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01872'} |
# Introduction and Summary
Energy conditions are indispensable in understanding classical and quantum gravity. The weakest but most commonly used of the standard energy conditions is the null energy condition (NEC), which states that \(T_{kk} \equiv T_{ab}k^{a}k^{b} \geq 0\) where \(T_{ab}\) is the stress tensor of the matter coupled to gravity and \(k^{a}\) is any null vector. It is sufficiently weak to be satisfied by familiar classical field theories, yet strong enough to prove the second law of black hole thermodynamics, singularity theorems, the chronology protection theorem, topological censorship, and other fundamental results. It also guarantees essential properties of holographic entanglement entropy in the context of gauge/gravity duality. On the other hand, it has long been known that the NEC is violated even in free quantum field theories. Several quantum replacements for the NEC have been suggested---such as the averaged null energy condition (ANEC) and "quantum inequalities" ---which involve integrating \(\braket{T_{kk}}\) over a region of spacetime, and others. In this paper we study the quantum null energy condition (QNEC), \[\begin{aligned}
\label{QNECintro} \braket{T_{kk}} \geq \frac{1}{2\pi} S''~, \end{aligned}\] which places a bound on the renormalized \(\braket{T_{kk}}\) at a point \(p\) in terms of a particular second derivative of the renormalized von Neumann entropy of a region touching \(p\) with respect to deformations of the region at \(p\).[^1] In [\[QNECintro\]](#QNECintro){reference-type="eqref" reference="QNECintro"} we have set \(\hbar =1\). Although shown in [\[QNECintro\]](#QNECintro){reference-type="eqref" reference="QNECintro"}, we will often omit the expectation value brackets below. The conjecture of states that [\[QNECintro\]](#QNECintro){reference-type="eqref" reference="QNECintro"} holds when \(k^a\) generates a locally stationary horizon through \(p\); i.e., it generates a hypersurface-orthogonal null congruence with vanishing shear \(\sigma_{ab}\) at \(p\) and expansion \(\theta\) at \(p\) vanishing to second order along the generator (\(\sigma_{ab}|_p = \theta|_p = \dot{\theta}|_p = 0\)). Below, we restrict to backgrounds satisfying the null convergence condition \(R_{ab}k^a k^b \ge 0\), so the Raychaudhuri equation \[\label{eq:Ray} \dot{\theta} =-\frac{\theta^2}{d-2}-\sigma^{ab} \sigma_{ab}-R_{ab} k^a k^b\] then requires \(R_{ab}k^a k^b = 0\) at \(p\). This conjecture was motivated in by taking a non-gravitating (\(G \rightarrow 0\)) limit of a "quantum focussing conjecture\" (QFC), which was in turn motivated by the generalized second law (GSL) of thermodynamics and the proposed covariant entropy bound. Such conjectures suffice to preserve the most fundamental of the above results even in the presence of quantum corrections . For example, although quantum corrections allow the formation of traversable wormholes, the GSL severely limits their utility. The QNEC has been proven for deformations along bifurcate Killing horizons in free bosonic theories using the techniques of null quantization, and it was also shown to hold for holographic theories formulated in flat space in. In the holographic case, the Ryu-Takayanagi-Hubeny-Rangamani formula was used to translate the QNEC (applied in the boundary theory) at leading order in \(1/N\) into a statement about how boundary-anchored extremal surfaces in AdS move when the anchoring region is deformed. The relevant condition is that when the boundary region is deformed within its domain of dependence, the corresponding extremal surface should move in a spacelike way, at least near the boundary. This condition---called "entanglement wedge nesting" in ---is automatically true assuming entanglement wedge reconstruction, and can also be proven directly from the NEC applied in the bulk. It was shown in that the QNEC continues to hold at all orders in \(1/N\), assuming the entanglement wedge nesting property and the quantum extremal surfaces prescription of (building on ). And as we note in section [3](#sec:HologProofs){reference-type="ref" reference="sec:HologProofs"} below, the Koeller-Leichenauer holographic argument admits a straightforward extension to arbitrary backgrounds. However, the quantities to which the resulting inequality applies are naturally divergent. One might expect that the inequality takes the form of a QNEC for the "bare" quantities that have not been fully renormalized. Though even this remains to be shown, it is nevertheless of central interest to understand when local counter-terms contribute to [\[QNECintro\]](#QNECintro){reference-type="eqref" reference="QNECintro"}--or, more specifically, when they contribute to the difference between the left-and right-hand sides. In such cases, a QNEC for renormalized quantities would also depend on the choice of renormalization scheme, as such choices induce finite shifts in naively-divergent couplings. We therefore refer to this phenomenon as scheme-dependence. A result of shows the QNEC to be scheme-independent when the null congruence \(N\) lies on a bifurcate Killing horizon (see and for precursors in special cases), but the more general setting is studied in section [2](#sec:schemedep){reference-type="ref" reference="sec:schemedep"} below. For \(d\le 3\) we find that this extends[^2] to locally stationary null congruences \(N\), and for \(d=4,5\) it holds when some additional derivatives of the shear \(\sigma_{ab}\) and expansion \(\theta\) also vanish at \(p\). For \(d \ge 6\) we show that scheme-independence generally fails even for weakly isolated horizons satisfying the dominant energy condition. The qualitative difference between the above cases is that in \(d\leq5\), finite counter-terms can only depend algebraically on the Riemann tensor by dimensional analysis, while in \(d\geq6\) derivatives of the Riemann tensor become allowed. In Sec. [2.2.3](#scheme-indep){reference-type="ref" reference="scheme-indep"} we show explicitly that the QNEC fails to be invariant under the addition of the simplest possible such counter-term, \(\int d^{6}x \sqrt{-g}(\nabla_{a}R)(\nabla^{a}R)\). In all cases where we find the QNEC to be scheme-independent, we show in section [3](#sec:HologProofs){reference-type="ref" reference="sec:HologProofs"} that a renormalized QNEC can be proven for the universal sector of holographic theories using the method of. This in particular establishes the QNEC for holographic theories on arbitrary backgrounds when the null congruence \(N\) lies on a bifurcate Killing horizon. We close with some final discussion in section [\[sec:discussion\]](#sec:discussion){reference-type="ref" reference="sec:discussion"}. In particular, we note that a corollary to our work is a general proof of the QFC and GSL for holographic quantum field theories in \(d\le 3\) at leading order in both \(1/N\) and the coupling \(G\) to gravity[^3], and for these theories on weakly-isolated horizons in \(d\leq5\). This is the first proof of the GSL in the semi-classical regime which does not require the quantum fields to be perturbations to a Killing horizon.
# Scheme-(in)dependence of the QNEC {#sec:schemedep}
As discussed above, a crucial question is whether local counter-terms affect [\[QNECintro\]](#QNECintro){reference-type="eqref" reference="QNECintro"}. To answer this question, it is useful to be more precise about how the various terms in [\[QNECintro\]](#QNECintro){reference-type="eqref" reference="QNECintro"} are to be computed. We briefly review such recipes in section [2.1](#subsec:prelim){reference-type="ref" reference="subsec:prelim"} and then consider the effect of local counter-terms in [2.2](#subsec:dotheycancel){reference-type="ref" reference="subsec:dotheycancel"}.
## Preliminaries {#subsec:prelim}
We consider states \(|\Psi\rangle\) that are pure on a sufficiently enlarged spacetime (with metric \(g_{ab}\)) and which are defined by a path integral with arbitrary operator insertions \({\cal O}[\Phi]\) and sources (included in the action \(I\)): \[\begin{aligned}
\label{PSIPI} \ket{\Psi} = \int_{\tau =-\infty}^{\tau = 0} [D\Phi] \mathcal{O}[\Phi] e^{-I[\Phi,g_{ab}]} \end{aligned}\] We use Euclidean notation for familiarity, though the integration contour may also include Lorentzian or complex pieces of spacetime (as in e.g. the discussion of entropy in ). The partition function \(Z=\text{Tr}\left[ \left| \Psi \right\rangle \left\langle \Psi \right| \right]\) for such a state is computed by sewing together two copies of [\[PSIPI\]](#PSIPI){reference-type="eqref" reference="PSIPI"} to form the path integral \[\begin{aligned}
\label{eq:Z} Z[g_{ab}] = \int [D\Phi] \mathcal{O}[\Phi] e^{-I[\Phi,g_{ab}]}. \end{aligned}\] As usual, [\[eq:Z\]](#eq:Z){reference-type="eqref" reference="eq:Z"} is a functional of the background geometry \(g_{ab}\). We take the definitions of \([D\Phi]\) and \(I[\Phi,g_{ab}]\) to include appropriate renormalizations to make \(Z[g_{ab}]\) well-defined. The renormalized effective action \(W\) is defined by \[\begin{aligned}
Z[g_{ab}] = e^{-W[g_{ab}]}. \end{aligned}\] Both \(T_{kk}\) and \(S''\) are to be computed from \(W[g_{ab}]\). The expectation value of the renormalized gravitational stress tensor in \(\ket{\Psi}\) is defined by \[\begin{aligned}
\label{Tdef} \braket{T_{ab}} \equiv-\frac{2}{\sqrt{-g}} \frac{\delta W}{\delta g^{ab}}. \end{aligned}\] Given a region \({\cal R}\) with boundary \(\Sigma = \partial {\cal R}\), we take the renormalized \(S\) to be computed from \(W[g_{ab}]\) via the replica trick, as the response of \(W[g_{ab}]\) to a conical singularity at \(\Sigma\) (henceforth called the entangling surface): \[\begin{aligned}
\label{Sdef} S &=-{\rm Tr} \rho \log \rho = (1-\partial_{n}) \log {\rm Tr} \rho^{n}. \end{aligned}\] The density matrix can be written in terms of the path integral, \(\rho^{n} = \frac{Z[g_{(n)ab}]}{Z^{n}[g_{(1)ab}]}\), where \(g_{(n)ab}\) denotes the geometry with \(n\) replicas of the original geometry, glued together at the entangling surface. Thus \(S\) can be expressed in terms of \(W[g_{(n)ab}]\) as \[\begin{aligned}
S = W_{(1)}-\partial_{n} W_{(n)}\big|_{n=1}, \end{aligned}\] with \(W_{(n)} \equiv W[g_{(n)ab}]\). It then remains to compute \(S''\). Given a null congruence \(N\) orthogonal to \(\Sigma\) (by convention taken outgoing relative to the region \({\cal R}\)), we may vary \(\Sigma\) by displacing it along the null generators of \(N\). This defines an associated deformation of \({\cal R}\), and thus a change of \(S\). The quantity \(S''\) is well-defined at \(p\) when the second such variation takes the form \[\label{eq:Spp} \frac{\delta }{\delta \Sigma \left(y_1\right)} \left[\frac{1}{\sqrt{h}} \frac{\delta }{\delta \Sigma \left(y_p\right)}S \right] = \sqrt{h} \ S'' (y_p) \delta(y_1-y_p) + f(y_p,y_1)\] for smooth functions \(S''(y_p), f(y_p,y_1)\). Here \(y\) labels the generators of the null congruence \(N\), \(y_p\) is the null generator through \(p\), and \(h\) denotes the determinant of the metric on \(\Sigma\) in the \(y\)-coordinate system.[^4]
## The QNEC and local counter-terms {#subsec:dotheycancel}
The renormalized effective action \(W\) depends on the choice of renormalization scheme, though any two schemes will differ only by adding finite local counter-terms; i.e., by the addition to \(W\) of integrals of marginal or relevant operators built from background curvatures of \(g_{ab}\) and/or matter fields. Such terms might in principle affect either side of [\[QNECintro\]](#QNECintro){reference-type="eqref" reference="QNECintro"}. Below, we calculate the net effect on the QNEC quantity \[\label{eq:Q} Q := T_{ab}k^ak^b-\frac{1}{2\pi}S''\] at points \(p\) where \(k^c\) generates a locally stationary null congruence in a background satisfying the null convergence condition \(R_{ab} k^a k^b \ge 0\). In particular, as noted in the introduction we must have \[\label{Rkkvanishes} (R_{ab} k^a k^b)|_p=0\] at all such points. We consider a theory that approaches a (unitary) conformal fixed point in the UV. The possible terms thus depend on the spacetime dimension \(d\). We will assume in all cases that there are no scalar operators saturating the unitarity bound \(\Delta =\frac{d-2}{2}\), so the addition to \(W\) of kinetic terms like \(\int d^dx\sqrt{-g} (\partial \phi)^2\) are not allowed. For simplicity, for \(d=2\) we also neglect conserved currents as in this case they would require special treatment. Apart from this one case, in the absence of non-metric sources combining covariance with unitarity bounds forbids the appearance of terms in \(W\) involving CFT operators with spin \(j \ge 1\). For later use in section [3.5](#sec:killingproof){reference-type="ref" reference="sec:killingproof"} we note that, using an argument like that in footnote [\[Fufoot\]](#Fufoot){reference-type="ref" reference="Fufoot"} below, a result of thus shows that the QNEC is scheme-independent on any bifurcate Killing horizon.
### \(d\le 3\)
For \(d\le 3\), the terms one may add to \(W\) are only \[\begin{aligned}
\label{eq:dl3terms} \int d^dx \sqrt{-g} \phi_1\text{, }\int d^dx \sqrt{-g} \,R \phi_2, \end{aligned}\] for scalar operators \(\phi_1,\phi_2\) of dimensions \(\Delta_1,\Delta_2\) with \(\Delta_1 \le d\) and \(\Delta_2 \le d-2\). Note that this includes the case \(\phi_2 = \mathds{1}\). The contribution of the first term to \(T_{ab}\) is proportional to \(g_{ab}\) and thus vanishes when contracted with \(k^a k^b\) for null \(k^c\). Its explicit contribution to \(S\) also vanishes, so it does not affect the QNEC. For terms of the second form in [\[eq:dl3terms\]](#eq:dl3terms){reference-type="eqref" reference="eq:dl3terms"} one finds \[\begin{aligned}
\Delta S &=& 4 \pi \int_{C_B} d^{d-2}y \sqrt{h} \phi_2, \\ \Delta T_{ab} &=& 2 \left(\nabla_a \nabla_b-g_{ab} \nabla^2\right) \phi_2 -2 \left({R}_{ab}-\frac{1}{2}g_{ab} {R} \right) \phi_2. \end{aligned}\] We use equation [\[eq:Spp\]](#eq:Spp){reference-type="eqref" reference="eq:Spp"} to calculate the derivative of \(\Delta S\). The first derivative of \(\Delta S\) is \(\Delta S'\equiv \frac{1}{\sqrt{h}}\frac{\delta }{\delta \Sigma \left(y_p\right)}\Delta S=4\pi \left(\theta \phi _2+\dot \phi _2\right)\), where the definition of the expansion \(\theta :=\frac{1}{\sqrt{h}}\frac{\delta \sqrt{h}}{\delta \Sigma \left(y_p\right)}\) (equation (1.1) of ) is used. The second derivative of the entropy is then \(\Delta S''=4\pi \left(\dot \theta \phi _2+\theta \dot \phi _2+\ddot \phi_2\right)\). Since \(k^a\) is null, the \(kk\)-component of \(\Delta T_{ab}\) is \(\Delta T_{kk}=2\ddot \phi _2-2R_{kk}\phi _2=2\ddot \phi _2+2\phi _2\left(\dot \theta +\frac{\theta ^2}{d-2}+\sigma_{ab}\sigma^{ab}\right)\), where the Raychaudhuri equation [\[eq:Ray\]](#eq:Ray){reference-type="eqref" reference="eq:Ray"} is used. Thus, the change of the quantity [\[eq:Q\]](#eq:Q){reference-type="eqref" reference="eq:Q"} is \[\Delta Q := \Delta T_{kk}-\frac{1}{2\pi }\Delta S'' = 2 \phi_2 \left( \frac{\theta^2}{d-2} + \sigma_{ab} \sigma^{ab}\right)-2\theta \dot{\phi_2}.\] Both terms vanish on a locally stationary horizon, and in fact \(\sigma_{ab}\) vanishes identically for \(d=3\). So under the above conditions the QNEC is scheme-independent for \(d \le 3\). In fact, we see that it is really only necessary to impose \(\theta=0\).
### \(d=4,5\) {#SI45}
Increasing \(d\) leads to additional terms. The allowed terms for \(d=4,5\) are those in [\[eq:dl3terms\]](#eq:dl3terms){reference-type="eqref" reference="eq:dl3terms"} together with \[\begin{aligned}
\label{eq:d45} \int d^dx \sqrt{-g} \,R_{ab}R^{ab}\text{, }\int d^dx\sqrt{-g} R_{abcd}R^{abcd}. \end{aligned}\] Since scalar operators \(\phi\) have dimensions larger than \(\frac{d-2}{2} \ge 1\), for \(d=4,5\) they cannot be inserted into the terms [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"}. Note that terms like \(\int d^dx \sqrt{-g} R^2\) can be written as the second term in [\[eq:dl3terms\]](#eq:dl3terms){reference-type="eqref" reference="eq:dl3terms"} by taking \(\phi_2\) to involve \(R\), so such terms were already considered above. The contributions of [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"} to the QNEC quantity \(Q\) are complicated and do not appear to vanish at the desired points \(p\). Indeed, for \(d\ge 5\) the results of show that the particular combination of the terms in [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"} with \(\int d^dx \sqrt{-g} R^2\) that defines the Gauss-Bonnet term contributes \(\Delta Q \neq 0\) even on Ricci-flat backgrounds[^5], though the Gauss-Bonnet contribution to \(Q\) vanishes for \(d=4\). The fact that terms in [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"} are four-derivative counter-terms suggests that their contributions to \(\Delta T_{ab} k^a k^b\) and \(\Delta S''\) contain fourth derivatives of the horizon generator \(k^a\). It is thus natural to ask if we can force \(\Delta Q=0\) for [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"} by setting the first, second, and third derivatives of the expansion \(\theta\) and shear \(\sigma_{\alpha \beta}\) to zero, where here and from now on, we use indices \(\alpha, \beta, \ldots\) to indicate coordinates on \(\Sigma\). It turns out that this is the case. For \(d=4,5\) we impose the conditions \[\label{eq:highD} \begin{aligned} \theta |_p&=(\mathscr{D}_a\theta )|_p=(\mathscr{D}_b\mathscr{D}_a\theta )|_p=(\mathscr{D}_c\mathscr{D}_b\mathscr{D}_a\theta )|_p=0, \\ \sigma _{\alpha \beta }|_p&=(\mathscr{D}_a\sigma _{\alpha \beta })|_p=(\mathscr{D}_b\mathscr{D}_a\sigma _{\alpha \beta })|_p=(\mathscr{D}_c\mathscr{D}_b\mathscr{D}_a\sigma _{\alpha \beta })|_p=0, \end{aligned}\] where \(\mathscr{D}_a\) is the covariant derivative along the congruence \(N\). We will show below that (when combined with a positive energy condition) these requirements suffice to show \(\Delta Q =0\), though the question remains open which conditions are precisely necessary. A final condition we impose is that the background solve the Einstein equations with a source respecting the dominant energy condition (DEC) up to a term proportional to the metric.[^6] This is equivalent to requiring \(R_{ab}\) to be of the form \[\label{gDEC} R_{ab}=R_{ab}^\text{(DEC)}+\alpha g_{ab},\] for some scalar field \(\alpha\), where for any future-pointing causal (either timelike or null) vector field \(v^a\), the vector field \(-R^a_{~b}{}^\text{(DEC)}v^b\) must also be both future-pointing and causal. A short argument (see appendix [\[appendix:identities\]](#appendix:identities){reference-type="ref" reference="appendix:identities"}) using [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"} then shows that on the null generator through \(p\) we have \[\begin{aligned}
\label{Rak} R_{ab}k^{b} = fk_{a} + O(\lambda^3), \end{aligned}\] for some scalar function \(f\) and that \[\begin{aligned}
\label{Rkk} R_{abcd}k^{b} k^{d} = \zeta k_{a} k_{c} + O(\lambda^3), \end{aligned}\] for some scalar function \(\zeta\). Since [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"} implies that equation [\[nablak\]](#nablak){reference-type="eqref" reference="nablak"} holds at point \(p\), the contracted Bianchi identity implies that equation [\[eq:WIHBianchi\]](#eq:WIHBianchi){reference-type="eqref" reference="eq:WIHBianchi"} holds at point \(p\), namely \[\begin{aligned}
\label{eq:Bianchi} (k^a \partial_{a} f)|_p = (\frac{1}{2} k^a \partial_{a}R)|_p = (-k^a \partial_{a} \zeta)|_p. \end{aligned}\] The expressions for \(\Delta S\) for the counter-terms [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"} can be found in and the expressions for \(\Delta T_{ab}\) for these counter-terms can be calculated by using the definition [\[Tdef\]](#Tdef){reference-type="eqref" reference="Tdef"}. These expressions are simplified greatly by using conditions [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"} and [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"}. After such simplifications it is straightforward (see appendix [\[appendix:schemeindependence\]](#appendix:schemeindependence){reference-type="ref" reference="appendix:schemeindependence"}) to compute \(\Delta S''\) and, as shown in table [1](#tb:4dcounterterm){reference-type="ref" reference="tb:4dcounterterm"}, the above conditions suffice to force \(\Delta Q=0\) for both terms [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"} in all dimensions. In table [1](#tb:4dcounterterm){reference-type="ref" reference="tb:4dcounterterm"} we have used the notation \(\partial_k := k^a \partial_a\), which we continue to use elsewhere below. As a final comment, we note that a careful analysis of the calculation shows that although for \(d=5\) we require the full list of conditions [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"} show \(\Delta Q=0\), for \(d=4\) it suffices to use only a subset of the conditions. The reason is that for \(d=4\) we may choose to study the Gauss-Bonnet term \(R_{abcd} R^{abcd}-4 R_{ab}R^{ab} + R^2\) (instead of \(R_{abcd} R^{abcd}\)). This term is topological and so contributes to neither \(T_{ab}\) nor \(S''\), and to guarantee \(\Delta Q=0\) for the only remaining counter-term \(R_{ab}R^{ab}\) we need only [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"} and conditions on the first line of [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"}.
### \(d \ge 6\) {#scheme-indep}
In six dimensions, six-derivative counter-terms become allowed. Based on the results above, one might hope to maintain scheme-independence of the QNEC in this case by requiring even more derivatives of the extrinsic curvature to vanish. However, we now show that for \(d\ge 6\) the contribution \(\Delta Q\) is generally non-zero even on weakly isolated horizons (where \(\theta, \sigma_{\alpha \beta}\) vanish identically on \(N\) ) in backgrounds satisfying [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"}. Since all derivatives of \(\theta, \sigma_{\alpha\beta}\) vanish, the results [\[Rak\]](#Rak){reference-type="eqref" reference="Rak"}, [\[Rkk\]](#Rkk){reference-type="eqref" reference="Rkk"}, and [\[eq:Bianchi\]](#eq:Bianchi){reference-type="eqref" reference="eq:Bianchi"} now exactly hold on a finite neighborhood of point \(p\) on the horizon. Furthermore, we show in appendix [\[appendix:identities\]](#appendix:identities){reference-type="ref" reference="appendix:identities"} that on weakly isolated horizons the Riemann tensor \(R_{abck}\) can be written as \(R_{abck}=k_c\tilde{A}_{ab}+k_{[a}\tilde{B}_{b]c}\), where \(\tilde{A}_{ab}\) is antisymmetric and satisfies \(k^a\tilde{A}_{ab}\propto k_b\) and \(\tilde{B}_{ab}\) satisfies \(k^a\tilde{B}_{ab}\propto k_b\) and \(k^b\tilde{B}_{ab}\propto k_a\). This allows one to write down additional relations also listed in appendix [\[appendix:identities\]](#appendix:identities){reference-type="ref" reference="appendix:identities"}. Together, they allow one to show the QNEC to be unchanged by adding six-derivative counter-terms built from polynomial contractions of the Riemann tensor. The computations are presented in appendix [\[appendix:schemeindependence\]](#appendix:schemeindependence){reference-type="ref" reference="appendix:schemeindependence"} and the results are summarized in table [2](#tb:6dcounterterm){reference-type="ref" reference="tb:6dcounterterm"}.
However, we will shortly see that scheme-independence of the QNEC can fail even on weakly isolated horizons for counter-tems that contain derivatives of the Riemann tensor. There are four such counter-terms, \((\nabla _aR)(\nabla ^aR)\), \((\nabla _aR_{bc})(\nabla ^aR^{bc})\), \((\nabla _eR_{abcd})(\nabla ^eR^{abcd})\), and \((\nabla _aR_{bc})(\nabla ^bR^{ac})\). Neglecting total derivatives of the action, these counter-terms are not linearly independent and one can write the last two previous counter-terms in terms of the other ten. We will show this failure for the term \((\nabla _aR)(\nabla ^aR)\). From we have \[\begin{aligned} \Delta T_{kk}&= k^ak^b \Delta T_{ab}\\ &=-4k^ak^b\nabla _a\nabla _b(\nabla _e\nabla ^eR)-2(k^a\nabla _aR)(k^b\nabla _bR)\\ &=-4\partial _k^2(\nabla _a\nabla ^aR)-2(\partial _kR)^2. \end{aligned}\] The formula used to calculate the change in the entropy is given in, where it was written as \[\label{eq:DeltaSfor6d} \Delta S = \Delta S_{\text{G-Wald}} + \Delta S_{\text{Anomaly}}.\] On weakly isolated horizons, the "generalized Wald entropy" term in [\[eq:DeltaSfor6d\]](#eq:DeltaSfor6d){reference-type="eqref" reference="eq:DeltaSfor6d"} reduces to the ordinary Wald entropy and is given by \[\Delta S_{\text{G-Wald}}=\Delta S_\text{Wald}=-2\pi \times (-2\nabla _a\nabla ^aR)\times (-2)=-8\pi \nabla _a\nabla ^aR.\] On weakly isolated horizons, by applying equation (A.12) of, one finds that the "entropy anomaly" term in [\[eq:DeltaSfor6d\]](#eq:DeltaSfor6d){reference-type="eqref" reference="eq:DeltaSfor6d"} vanishes for counter-term \((\nabla _aR)(\nabla ^aR)\). Thus, \[\Delta S_{\text{Anomaly}}=0.\] Finally, we obtain \[\label{eq:DQ6d} \Delta Q=\Delta T_{kk}-\frac{1}{2\pi }\Delta S''=-2(\partial _kR)^2.\] This quantity does not vanish generally, although it vanishes on Ricci-flat background. It is not hard to find an explicit geometry in which this counter-term spoils the scheme-independence of the QNEC. Consider the spacetime metric \[ds^2=-dudv-dvdu-cu^2v^2du^2+\sum\limits_{\alpha }{{{\left( d{{y}_{\alpha }} \right)}^{2}}},\] where \(c\) is a positive constant which is not assumed to be small. In this spacetime, there is a non-expanding null surface \(v=0\). Its Ricci tensor is \[\label{eq:exRab} R_{ab}=c^2u^4v^2(du)_a(du)_b+cu^2(du)_a(dv)_b+cu^2(dv)_a(du)_b =-cu^2 g_{\perp ab},\] where \(g_{\perp ~b}^{~a}\) is the projector onto the \(u\)-\(v\) plane. Since this plane is timelike, for any future-pointing causal vector \(v^b\) the vector \(-R^a_{~b} v^b = cu^2 g^{~a}_{\perp ~b} v^b\) is again future-pointing and causal and the Ricci tensor [\[eq:exRab\]](#eq:exRab){reference-type="eqref" reference="eq:exRab"} satisfies [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"} with \(\alpha =0\). Thus the plane \(v=0\) is a weakly isolated horizon. But from [\[eq:exRab\]](#eq:exRab){reference-type="eqref" reference="eq:exRab"} we find scalar curvature \(R=-2cu^2\), hence \[\Delta Q=-32c^2u^2\neq 0\] and the QNEC fails to be scheme-independent.
# Holographic Proofs of the QNEC {#sec:HologProofs}
The QNEC was proven to hold in for leading-order holographic field theories on flat spacetimes. We review this derivation in section [3.1](#subsec:outline){reference-type="ref" reference="subsec:outline"} below and show that the argument admits a straightforward generalization to arbitrary curved backgrounds; i.e., to the case where the boundary of the asymptotically locally AdS bulk is arbitrary. However, the resulting inequality is generally divergent, and we expect it to yield a finite renormalized QNEC only in the contexts where the QNEC is scheme-independent. For the scheme-independent cases described in section [2](#sec:schemedep){reference-type="ref" reference="sec:schemedep"}, we will indeed be able to derive such a finite renormalized QNEC below. Since all the proofs in this section are provided in the context of AdS/CFT correspondence, we change our index notation to make it more suitable for this context. In this section, we use indices \(\mu ,\nu ,\ldots\) to indicate the \(d+1\) coordinates on the bulk spacetime and use indices \(i,j,\ldots\) to indicate the \(d\) coordinates on the boundary spacetime.
## Outline of holographic proofs {#subsec:outline}
The central idea of is to reformulate both \(T_{kk}\) and \(S''\) in terms of quantities in the dual bulk asymptotically AdS spacetime, and to use a fact about extremal surfaces known as "entanglement wedge nesting" (EWN) to provide the desired inequality. To begin, consider two regions \(A,B\) on the asymptotically AdS boundary. Entanglement wedge nesting states that if these boundary regions are nested in the sense that \(D(B) \subseteq D(A)\) then their extremal surfaces \(e(A)\), \(e(B)\) must also be nested, i.e. everywhere spacelike related. Here \(D(A)\) is the domain of dependence of \(A\) in the asymptotically AdS boundary. Now consider a family of boundary regions \(A(\lambda)\) with entangling surfaces \(\partial A(\lambda)\), which differ by localized deformations along a single null generator \(k^{i}(y)\) of a null hypersurface shot out from the initial surface \(A(0)\). The derivatives of the entropy used in the QNEC are then derivatives with respect to this particular \(\lambda\), which we can take to be an affine parameter for \(k^{i}\). Consider the codimension-1 bulk surface \(\mathcal{M}\) foliated by the (smallest for each \(A(\lambda)\)) extremal surfaces \(e(\lambda) \equiv e(A(\lambda))\), \[\begin{aligned}
\mathcal{M}:= \cup_{\lambda} e(\lambda) \end{aligned}\] EWN for \(A(\lambda)\) implies that \(\mathcal{M}\) is a spacelike surface. The surface \(\mathcal{M}\) can be parametrized by \(\lambda\) and the \(d-1\) coordinates \(y^{\alpha} = \{z,y^{a}\}\) associated with each \(e(\lambda)\). The coordinate basis for the tangent space of \(\mathcal{M}\) then consists of \(\partial_{\lambda}X^{\mu}\) and the \(d-1\) coordinate basis tangent vectors of the extremal surfaces, \(\partial_{\alpha} X^{\mu}\). By EWN, each of these vectors has positive norm. The norm of \(\partial_{\lambda}X^{\mu}\) can be expanded in \(z\) near the boundary, and will involve the near-boundary expansion of both the metric and the extremal surface embedding functions. EWN implies in particular that as \(z\to 0\) the most dominant term in this expansion is positive. In it was shown that for locally-stationary surfaces (satisfying \(\theta |_p= \sigma _{ab}|_p= 0\) at a point \(p\)) in flat space one has[^7] \[\begin{aligned}
\label{flatchi^{2}} 0 \leq G_{\mu\nu} \partial_{\lambda}X^{\mu} \partial_{\lambda}X^{\nu} = \frac{16\pi G}{d} z^{d-2} \left( T_{kk}-\frac{1}{2\pi} S'' \right), \end{aligned}\] where the quantities on the right-hand side are have been renormalized.[^8] Thus for these surfaces in flat space, EWN implies the renormalized QNEC. As indicated in equation [\[flatchi\^{2}\]](#flatchi^{2}){reference-type="eqref" reference="flatchi^{2}"}, the renormalized quantities in the QNEC appear in the EWN inequality at \(O(z^{d-2})\). So in order to derive the renormalized QNEC, the terms in the \(z\)-expansion of \((\partial_{\lambda}X^{\mu})^{2}\) must vanish at all lower orders. This condition provides restrictions on the surfaces and space-times. In flat space, it is sufficient to have local stationarity, i.e. \(\theta |_p=0, \sigma_{\alpha \beta}|_p=0\) at a point \(p\). More generally the condition may be more complicated. To compute [\[flatchi\^{2}\]](#flatchi^{2}){reference-type="eqref" reference="flatchi^{2}"} explicitly, we set \(\ell_{\text{AdS}}=1\) and use Fefferman-Graham-style coordinates to introduce the near-boundary expansion of the bulk metric \(G_{\mu \nu}\): \[\begin{aligned} G_{zz}(x^{\mu}) &= \frac{1}{z^{2}}\text{, }G_{zi} = 0,\\ \label{Gexpn} G_{ij}(x^{\mu}) &\equiv \frac{1}{z^{2}} g_{ij} = \frac{1}{z^{2}} \left( g_{(0)ij} + g_{(2)ij} + \ldots + g_{(dl)ij} + \bar{g}_{(d)ij} + \frac{16\pi G}{d} z^{d}T_{ij} \right), \end{aligned}\] and the embedding functions of the extremal surfaces \[\label{Xexpn} \begin{aligned} X^{z}(y^{\alpha}) &= z, \\ X^{i}(y^{\alpha}) &= X^{i}_{(0)} + X^{i}_{(2)} + \ldots + X^{i}_{(dl)}-\frac{4G}{d} z^{d} g_{(0)}^{ij} S'_{i} + \bar{X}^{i}_{(d)}, \end{aligned}\] where \[\begin{aligned}
\label{S'idef} S'_{i} \equiv \frac{1}{\sqrt{h_{(0)}}} \frac{\delta S_{\rm ren}}{\delta X^{i}_{(0)}} \end{aligned}\] is the renormalized entropy directional derivative per unit area and subscripts denote powers of \(z\), e.g. \(X^{i}_{(m)}\) is \(O(z^{m})\), while \(g_{(dl)ij}\) is a log term of order \(z^{d}\log{z}\). As a result, \(\partial _\lambda X^\mu\) involves \(S'' = \partial_\lambda(k^i S_i{}')\). The terms \(\bar{g}_{(d)ij}\) and \(\bar{X}^{i}_{(d)}\) refer to the geometric parts of the \(O(z^{d-2})\) and \(O(z^{d})\) parts of the metric and embedding functions, respectively. For more details, see for the extremal surface expansion, and e.g. for the metric. The key point is that \(S' \equiv k_{(0)}^{i}S'_{i}\) and \(T_{ij}k^{i}_{(0)}k^{i}_{(0)}\) appear at \(O(z^{d-2})\). This is why plugging these expansions into \((\partial_{\lambda}X^{\mu})^{2}\) gives the QNEC at \(O(z^{d-2})\), as in [\[flatchi\^{2}\]](#flatchi^{2}){reference-type="eqref" reference="flatchi^{2}"}. That the CFT stress tensor appears at \(O(z^{d-2})\) in the near-boundary expansion for the metric is well known. We now derive the appearance of \(S'_{i}\) in equation [\[Xexpn\]](#Xexpn){reference-type="eqref" reference="Xexpn"}. In Einstein-Hilbert gravity, the HRT prescription is \[\begin{aligned}
\label{eq:HRTaction} S(\lambda) = \frac{A(e(\lambda))}{4G} = \frac{1}{4G} \int dz d^{d-2}y \sqrt{H[X]}, \end{aligned}\] where \(H[X]\) is the determinant of the induced metric on \(e(\lambda)\) written as a functional of \(X\). Varying the on-shell area functional with respect to \(X^{i}\) gives a boundary term evaluated at a cutoff surface \(z = \rm const\). This produces the regulated entropy variation \[\begin{aligned}
\label{deltaS} \frac{1}{\sqrt{h}}\frac{\delta S_{\rm reg}}{\delta X^{i}} =-\frac{1}{4G} z^{1-d} \frac{g_{ij} \partial_{z} X^{j}}{\sqrt{1+g_{nm}\partial_{z}X^{n}\partial_{z}X^{m}}} \Bigg|_{z=\rm const}, \end{aligned}\] where \(h\) is the determinant of the induced metric. Everything in equation [\[deltaS\]](#deltaS){reference-type="eqref" reference="deltaS"} (including \(\delta X^{i}\)) is to be expanded in \(z\) and evaluated at a cutoff surface at \(z=\rm const\). In general, there will be terms that diverge as \(z \to 0\). The entropy can be renormalized using in a manner similar to that used for the on-shell action (see e.g. ). Indeed, the two are intimately related. Adding local, geometrical, covariant counter-terms gives the renormalized entropy via \[\begin{aligned}
S_{\rm ren} = \lim_{z\to 0} \,\left(S_{\rm reg} + S_{\rm ct}\right). \end{aligned}\] With an arbitrary choice of renormalization scheme, \(S_{\rm ct}\) can contain finite counter-terms in addition to those required to cancel divergences. Expanding [\[deltaS\]](#deltaS){reference-type="eqref" reference="deltaS"} and removing the divergences, we have the finite renormalized entropy variation given in general by \[\begin{aligned}
\label{Srenexpn} \frac{1}{\sqrt{h_{(0)}}}\frac{\delta S_{\rm ren}}{\delta X^{i}_{(0)}} =-\frac{d}{4G z^{d}} g_{(0)ij} X^{j}_{(d)} + \ldots \end{aligned}\] where \(h_{(0)}\) is the determinant of the metric induced on \(\Sigma\) by \(g_{(0)ij}\), \(X^{j}_{(d)}\) denotes the \(O(z^{d})\) part of \(X^{j}\), and the "\(\ldots\)" denotes finite, local, geometric terms, which include both finite contributions to equation [\[deltaS\]](#deltaS){reference-type="eqref" reference="deltaS"} from products of lower-order terms in the embedding functions, as well as possible finite counter-terms from \(S_{\rm ct}\). Re-arranging for \(X^{i}_{(d)}\) gives \[\begin{aligned}
\label{X^{i}_{d}} X^{i}_{(d)} =-\frac{4G}{d} z^{d} g_{(0)}^{ij} \frac{1}{\sqrt{h_{(0)}}} \frac{\delta S_{\rm ren}}{\delta X^{j}_{(0)}} + \bar{X}^{i}_{(d)} \end{aligned}\] In this expression, \(\bar{X}^{i}_{(d)}\) contains the contribution from the "\(\ldots\)" of equation [\[Srenexpn\]](#Srenexpn){reference-type="eqref" reference="Srenexpn"}. Plugging [\[X\^{i}\_{d}\]](#X^{i}_{d}){reference-type="eqref" reference="X^{i}_{d}"} into the expansion of \(X^{i}\) yields equation [\[Xexpn\]](#Xexpn){reference-type="eqref" reference="Xexpn"}, as promised. The geometric restrictions on the surface and geometry which guarantee that the lower-order terms in expression [\[flatchi\^{2}\]](#flatchi^{2}){reference-type="eqref" reference="flatchi^{2}"} vanish can in principle be determined by solving the relevant equations, though we will not carry out an exhaustive analysis here. We show in section [3.2](#subsec:Hd=3){reference-type="ref" reference="subsec:Hd=3"} below that the above argument leads to a finite renormalized scheme-independent QNEC for \(d\le 3\) at points \(p\) where the expansion \(\theta\) vanishes for the chosen null congruence \(N\). We then show in [3.3](#subsec:Hd=4){reference-type="ref" reference="subsec:Hd=4"} and [3.4](#subsec:Hd=5){reference-type="ref" reference="subsec:Hd=5"} below that for \(d=4,5\) it leads to a finite renormalized scheme-independent QNEC at points \(p\) where the chosen null congruence \(N\) satisfies the conditions on the first line of [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"} on backgrounds satisfying [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"}. From the results of section [2.2.2](#SI45){reference-type="ref" reference="SI45"} and the fact that--for Einstein-Hilbert gravity in the bulk--holographic renormalization requires only counter-terms that can be built from the Ricci tensor, it is no surprise that we do not require the full list of conditions [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"}. Finally, in section [3.5](#sec:killingproof){reference-type="ref" reference="sec:killingproof"} we provide a finite renormalized scheme-independent QNEC for holographic theories on Killing horizons in arbitrary backgrounds.
## Proof of the \(d\le3\) holographic QNEC {#subsec:Hd=3}
For \(d=3\) the asymptotic metric expansion [\[Gexpn\]](#Gexpn){reference-type="eqref" reference="Gexpn"} and the asymptotic embedding function expansion [\[Xexpn\]](#Xexpn){reference-type="eqref" reference="Xexpn"} take the form \[\begin{aligned}
\label{eq:expns3d} g_{ij}(x,z) &= g_{(0)ij} + g_{(2)ij} + g_{(3)ij} + \ldots, \\ X^i(x,z) &= X^i_{(0)} + X^i_{(2)} + X^i_{(3)} +\ldots. \end{aligned}\] The causal property of extremal surfaces then implies \[\begin{aligned} \label{eq:inequality3} 0 &\leq g_{ij}(\partial_{\lambda}X ^i)(\partial_{\lambda}X^j) \\ &= g_{(0)ij}k^{i}k^{j} + g_{(2)ij}k^{i}k^{j} + 2 g_{(0)ij} (\nabla_{\lambda} X_{(2)}^{i}) k^{j} + g_{(3)ij} k^ik^j + 2 g_{(0)ij} (\nabla_{\lambda}X^i_{(3)})k^j. \end{aligned}\] One can easily check that the Einstein equations and extremal surface equation at second order in \(z\) give \[\begin{aligned}
\label{g2X2} g_{(2)ij} = \frac{z^{2}}{d-2} \left( R_{ij}-\frac{1}{2(d-1)} R \, g_{(0)ij}\right)~, \quad X_{2}^{i} = \frac{1}{2(d-2)} z^{2} K^{i}~, \end{aligned}\] in terms of the (traced) extrinsic curvature \(K^i: = g_{(0)}^{jk} K^i_{~jk}\) of the (boundary) codimension-2 surface \(\partial A\) with conventions given by [\[eq:DefExtrCurv\]](#eq:DefExtrCurv){reference-type="eqref" reference="eq:DefExtrCurv"}. Since \(k^i\) is null, \(\dot{\theta} = \nabla_{\lambda}(g_{(0)ij} K^{i}k^{j})\) and that \(k^i\) satisfies the geodesic equation \(\nabla_\lambda k^i := k^j \nabla_j k^i = 0\), the terms on the second line of [\[eq:inequality3\]](#eq:inequality3){reference-type="eqref" reference="eq:inequality3"} combine to give \[\begin{aligned}
\frac{z^2}{d-2} \left( R_{ij} k^i k^j + \dot{\theta} \right) =-\frac{z^2}{d-2} \left( \frac{\theta^2}{d-2} + \sigma^{ij}\sigma_{ij}\right), \end{aligned}\] where in the last step we have used [\[eq:Ray\]](#eq:Ray){reference-type="eqref" reference="eq:Ray"}. Both terms vanish on a locally stationary horizon, and in fact \(\sigma_{ij}\) vanishes identically for \(d=3\). The terms on the third line of [\[eq:inequality3\]](#eq:inequality3){reference-type="eqref" reference="eq:inequality3"} then give the renormalized QNEC [\[flatchi\^{2}\]](#flatchi^{2}){reference-type="eqref" reference="flatchi^{2}"}. The \(d=2\) argument is identical in form. Though terms at order \(z^2\) are not divergent for \(z=2\), there are then divergent terms at order \(z^2 \log z\) which are structurally the same as the \(z^2\) terms for \(d=3\).
## Proof of the \(d=4\) holographic QNEC {#subsec:Hd=4}
For the case of four dimensional boundary, the asymptotic metric expansion [\[Gexpn\]](#Gexpn){reference-type="eqref" reference="Gexpn"} and the asymptotic embedding function expansion [\[Xexpn\]](#Xexpn){reference-type="eqref" reference="Xexpn"} take the form \[\begin{aligned}
\label{eq:expns} g_{ij}(x,z) &= g_{(0)ij} + g_{(2)ij} + g_{(4l)ij} + g_{(4)ij} + \ldots, \\ X^i(x,z) &= X^i_{(0)} + X^i_{(2)} + X^i_{(4l)} + X^i_{(4)} +\ldots. \end{aligned}\] The causal property of extremal surfaces provides us the inequality in QNEC. We have \[\begin{aligned} \label{eq:inequality} 0 &\leq g_{ij}(\partial_{\lambda}X ^i)(\partial_{\lambda}X^j) \\ &= g_{(0)ij}k^{i}k^{j} + g_{(2)ij}k^{i}k^{j} + 2 g_{(0)ij} (\nabla_{\lambda} X_{(2)}^{i}) k^{j} \\ &+ g_{(4l)ij} k^ik^j + 2 g_{(0)ij} (\nabla_{\lambda}X^i_{(4l)})k^j + g_{(4)ij} k^ik^j + 2 g_{(0)ij} (\nabla_{\lambda}X^i_{(4)})k^j. \end{aligned}\] Terms on the second line of [\[eq:inequality\]](#eq:inequality){reference-type="eqref" reference="eq:inequality"} vanish just as in section [3.2](#subsec:Hd=3){reference-type="ref" reference="subsec:Hd=3"}. From we have \[\label{eq:g4lkk} g_{(4l)ij}k^ik^j=-\frac{z^4\log z}{24}\partial _k^2R,\] \[\label{eq:g4kk} g_{(4)ij}k^ik^j = z^{4} \left(4\pi GT_{kk}+\frac{1}{32}\partial _k^2R\right),\] which may be used to calculate the third line of equation [\[eq:inequality\]](#eq:inequality){reference-type="eqref" reference="eq:inequality"}. Here and below we freely use equation [\[nablak\]](#nablak){reference-type="eqref" reference="nablak"} which follows from the first line of [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"}. We will first show that the log terms cancel each other, and then show that the \(O(z^{4})\) terms produce the QNEC. We introduce the standard notation \[\label{Aren} A_\text{ren} = A_\text{reg} + A_\text{ct}.\] The entropy counter-terms \(A_\text{ct}\) generically contain a finite part which we must extract. This comes from the requirement that the counter-terms are covariant functionals of the geometric quantities on the cutoff surface. The counter-term \(A_{\text{ct, }O\left(\log z\right)}\) which cancels the log divergence has an explicit \(\log z\), and consequently has no finite part. The finite part \(A_\text{ct, finite}\) of the counter-terms comes from the counter-term which cancels the leading area-law divergence. This is \[\begin{aligned}
A_\text{ct, A} &=-\frac{1}{d-2} \int_{z=\epsilon} d^{d-2}y\sqrt{\gamma } =-\frac{1}{2} \int_{z=\epsilon} d^{d-2}y\sqrt{\gamma ^{(0)}} \frac{1}{z^2}\left[ 1 + \frac{1}{2} g^{||ij}g_{(2)ij} + \ldots \right], \end{aligned}\] where \(\sqrt{\gamma }\) is the induced metric on the intersection of the HRT surface with the cutoff surface, and \(g^{\parallel ij}\) is the part of the boundary metric parallel to the entangling surface. Using the first equality of [\[eq:Bianchi\]](#eq:Bianchi){reference-type="eqref" reference="eq:Bianchi"}, the finite part of the counter-term can be written \[\begin{aligned} A_\text{ct, finite}&=-\frac{1}{4}\int d^{d-2}y\sqrt{\gamma ^{(0)}}\left(g^{ij}-g^{\perp ij}\right)\left(-\frac{1}{2}R_{ij}+\frac{1}{12}Rg_{ij}\right)\\ &= -\frac{1}{4}\int d^{d-2}y\sqrt{\gamma ^{(0)}}\left(f-\frac{R}{3}\right). \end{aligned}\] Thus we find \[A'_\text{ct, finite}=-\frac{1}{24}\partial _kR.\] From Eq. [\[deltaS\]](#deltaS){reference-type="eqref" reference="deltaS"} we have \[\label{A'reg} A'_\text{reg} =-\frac{1}{z^{d-1}} g_{ij} (\partial _{z}X^i) (\partial _{\lambda }X^j).\] Expanding the above equation and using equation [\[g2X2\]](#g2X2){reference-type="eqref" reference="g2X2"}, we find that the non-trivial contributions are at the same order as in [\[eq:inequality\]](#eq:inequality){reference-type="eqref" reference="eq:inequality"}, which are \[A'_\text{reg} =-\frac{1}{z^{3}} \left[ g_{(0)ij} (\nabla_{z} X^i_{(4l)}) k^j|_{z^3\log z}+g_{(0)ij} (\nabla_{z} X^i_{(4l)}) k^j|_{z^3} + g_{(0)ij} (\nabla_{z} X^i_{(4)}) k^j \right].\] The first term on right-hand side is the logarithmic divergence of the first derivative of the entropy, which can be computed as \[A'_{\text{reg, }O\left(\log z\right)}=-\frac{1}{2}(\log z)\partial _k\left(f-\frac{R}{3}\right)=-\frac{1}{12}(\log z)\partial _kR\] From this, we can infer that \[\begin{aligned}
g_{(0)ij}(\nabla _zX^i_{(4l)})k^j|_{z^3\log z}=&\frac{1}{12}z^3(\log z)\partial _kR, \\ g_{(0)ij}X^i_{(4l)}k^j=&\frac{1}{48}z^4(\log z)\partial _kR, \\ g_{(0)ij}(\nabla _zX^i_{(4l)})k^j|_{z^3}=&\frac{1}{48}z^3\partial _kR,\\ g_{(0)ij}(\nabla _{\lambda }X^i_{(4l)})k^j=&\frac{1}{48}z^4(\log z)\partial _k^2R. \end{aligned}\] As expected, the \(O\left(z^4\log z\right)\) terms in equation ([\[eq:inequality\]](#eq:inequality){reference-type="ref" reference="eq:inequality"}) cancel: \[g_{(4l)ij} k^ik^j + 2 g_{(0)ij} (\nabla_{\lambda}X^i_{(4l)})k^j=0.\] And since the rate of change of the renormalized area is \[\begin{aligned}
A'_\text{ren}= A'_\text{reg, finite}+A'_\text{ct, finite} = -\frac{1}{z^3}\left[\frac{1}{48}z^3\partial _kR+g_{(0)ij}(\partial _zX^i_{(4)})k^j\right]-\frac{1}{24}\partial _kR, \end{aligned}\] we obtain \[\begin{aligned}
g_{(0)ij}(\nabla _zX^i_{(4)})k^j= & z^3\left(-A'_\text{ren}-\frac{1}{16}\partial _kR\right),\\ g_{(0)ij}X^i_{(4)}k^j= & \frac{z^4}{4}\left(-A'_\text{ren}-\frac{1}{16}\partial _kR\right),\\ g_{(0)ij}(\nabla _\lambda X^i_{(4)})k^j= & z^4\left(-\frac{A''_\text{ren}}{4}-\frac{1}{64}\partial _k^2R\right). \label{g0X4k} \end{aligned}\] We now have all we need to evaluate the rest of the final line of equation [\[eq:inequality\]](#eq:inequality){reference-type="eqref" reference="eq:inequality"} and derive the QNEC. Plugging [\[g0X4k\]](#g0X4k){reference-type="eqref" reference="g0X4k"} and equation ([\[eq:g4kk\]](#eq:g4kk){reference-type="ref" reference="eq:g4kk"}) into equation ([\[eq:inequality\]](#eq:inequality){reference-type="ref" reference="eq:inequality"}), we obtain the QNEC \[0\leq T_{kk}-\frac{A''_\text{ren}}{8\pi G}.\]
## Proof of the \(d=5\) holographic QNEC {#subsec:Hd=5}
The \(d=5\) case is similar. We find \[\begin{aligned} \label{eq:inequality5d} 0 &\leq g_{ij}(\partial_{\lambda}X^i) (\partial_{\lambda}X^i) \\ &= g_{(0)ij}k^{i}k^{j} + g_{(2)ij}k^{i}k^{j} + 2 g_{(0)ij} (\nabla_{\lambda} X_{(2)}^{i}) k^{j} \\ &+ g_{(4)ij} k^ik^j + 2 g_{(0)ij} (\nabla_{\lambda}X^i_{(4)})k^j + g_{(5)ij} k^ik^j + 2 g_{(0)ij} (\nabla_{\lambda}X^i_{(5)})k^j, \end{aligned}\] with terms on the second line vanishing just as in section [3.2](#subsec:Hd=3){reference-type="ref" reference="subsec:Hd=3"}. As before, we will freely use equation [\[nablak\]](#nablak){reference-type="eqref" reference="nablak"} which follows from [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"}. For \(d=5\) we find \[\label{eq:g4kk5d} g_{(4)ij}k^ik^j=\frac{z^4}{32}\partial _k^2R,\] \[\label{eq:g5kk} g_{(5)ij}k^ik^j=z^5\frac{16\pi G}{5}T_{kk}.\] We again consider [\[Aren\]](#Aren){reference-type="eqref" reference="Aren"}. Because the boundary dimension is odd, the counter-term \[\begin{aligned}
A_\text{ct, A} &=-\frac{1}{d-2} \int_{z=\epsilon} d^{d-2}y\sqrt{\gamma } =-\frac{1}{3} \int_{z=\epsilon} d^{d-2}y\sqrt{\gamma ^{(0)}} \frac{1}{z^3}\left[ 1 + \frac{1}{2} g^{||ij}g_{(2)ij} + \ldots \right]. \end{aligned}\] contributes no finite part to the renormalized entropy; \[\begin{aligned}
A_\text{ct, finite} =0. \end{aligned}\] From Eq. [\[deltaS\]](#deltaS){reference-type="eqref" reference="deltaS"} we have \[\begin{aligned}
\label{Aprime} A'_\text{reg} =-\frac{1}{z^{d-1}} g_{ij} (\partial_{z}X^i) (\partial_{\lambda}X^j). \end{aligned}\] Expanding the above equation, we find that the non-trivial contributions are at the same order as above, which are \[\begin{aligned}
A'_{\rm reg} =-\frac{1}{z^{4}} g_{(0)ij} (\nabla_{z} X^i_{(4)}) k^j-\frac{1}{z^{4}} g_{(0)ij} (\nabla_{z} X^i_{(5)}) k^j. \end{aligned}\] The first term on right-hand side is a divergence which must be canceled by the counter-term and the second term on right-hand side is the finite part. We introduce the following notations: \[\begin{aligned}
A'_{\text{reg, }O(z^{-1})} =-\frac{1}{z^{4}} g_{(0)ij} (\nabla_{z} X^i_{(4)}) k^j, \\ A'_\text{reg, finite} =-\frac{1}{z^{4}} g_{(0)ij} (\nabla_{z} X^i_{(5)}) k^j. \end{aligned}\] Using [\[Rak\]](#Rak){reference-type="eqref" reference="Rak"}, the \(O(z^{-1})\) divergent part of the regulated area can be computed as \[\begin{aligned} A_{\text{reg, }O(z^{-1})}&=\int d^{d-2}y\sqrt{\gamma ^{(0)}}\frac{1}{2z}\left(\frac{1}{3}R_{ij}g^{\perp ij}-\frac{5}{24}R\right)\\ &=\int d^{d-2}y\sqrt{\gamma ^{(0)}}\frac{1}{2z}\left(\frac{2}{3}f-\frac{5}{24}R\right). \end{aligned}\] Together with the first equality in [\[eq:Bianchi\]](#eq:Bianchi){reference-type="eqref" reference="eq:Bianchi"}, this further implies \[A'_{\text{reg, }O(z^{-1})}=\frac{1}{16z}\partial _kR.\] Therefore we have \[\begin{aligned}
g_{(0)ij}(\nabla _z X^i_{(4)})k^j =-\frac{z^3}{16}\partial _kR,\\ g_{(0)ij}X^i_{(4)}k^j =-\frac{z^4}{64}\partial _kR,\\ g_{(0)ij}(\nabla _\lambda X^i_{(4)})k^j =-\frac{z^4}{64}\partial _k^2R. \label{g0X4kd=5} \end{aligned}\] Plugging equation [\[g0X4kd=5\]](#g0X4kd=5){reference-type="eqref" reference="g0X4kd=5"} and [\[eq:g4kk5d\]](#eq:g4kk5d){reference-type="eqref" reference="eq:g4kk5d"} into [\[eq:inequality5d\]](#eq:inequality5d){reference-type="eqref" reference="eq:inequality5d"}, we find as expected that the \(O(z^{4})\) terms cancel: \[\label{Oz4} g_{(4)ij} k^ik^j + 2 g_{(0)ij} (\nabla_{\lambda}X^i_{(4)})k^j = 0.\] Combining the above results yields \[\begin{aligned}
A'_\text{ren}= & A'_\text{reg, finite}+A'_\text{ct, finite} = -\frac{1}{z^4}g_{(0)ij}(\nabla _zX^i_{(5)})k^j+0, \end{aligned}\] which implies \[\begin{aligned}
\label{eq:renArea} g_{(0)ij}X^i_{(5)}k^j =-\frac{z^{5}}{5} A'_{\rm ren}. \end{aligned}\] Using ([\[eq:inequality5d\]](#eq:inequality5d){reference-type="ref" reference="eq:inequality5d"}), ([\[eq:g5kk\]](#eq:g5kk){reference-type="ref" reference="eq:g5kk"}), ([\[Oz4\]](#Oz4){reference-type="ref" reference="Oz4"}), ([\[eq:renArea\]](#eq:renArea){reference-type="ref" reference="eq:renArea"}), we thus obtain the renormalized QNEC \[\begin{aligned}
0 \leq T_{kk}-\frac{1}{8\pi G} A''_{\rm ren}. \end{aligned}\]
## Killing horizons {#sec:killingproof}
As noted in section [2](#sec:schemedep){reference-type="ref" reference="sec:schemedep"}, a result of implies the QNEC to be scheme-independent on any bifurcate Killing horizon. So for any \(d\) we would expect the holographic argument to yield a finite renormalized QNEC in this case as well. In particular, we consider a boundary metric \(g_{(0)ij}\) with a bifurcate Killing horizon \(H_{(0)}\) generated by the Killing vector \(\begin{align} \mathcal{L}_{ \end{align} We evaluate the QNEC for deformations generated by \)k^i \propto The critical fact is that, as explained above, the possible (finite or divergent) corrections to \eqref{flatchi^{2}} are of the form \(Z_{ij} k^i k^j\) for some smooth (covariant) geometric tensor \(Z_{ij}\) built from the spacetime metric \(g_{(0)ij}\), the extrinsic curvature \(K^i_{~jk}\), the projector \(h_{(0)i}{}^j\) onto \(\partial A\), and their derivatives. Since \( \section{Discussion} \label{sec:discussion} We have investigated the QNEC in curved space by analyzing the scheme-independence of the QNEC and its validity in holographic field theories. For \)d\le 3\(, for arbitrary backgroud metric we found that the QNEC \eqref{QNECintro} is naturally finite and independent of renormalization scheme for points \)p\( and null congruences \)N\( for which the expansion \)\theta\( vanishes at \)p\(. It is interesting that this condition is weaker than the local stationarity assumption (\)\theta |_p= \dot{\theta} |_p=0, \sigma_{ab}|_p=0\() under which the QNEC was previously proposed to hold, and it is in particular weaker than the conditions under which it can be derived from the quantum focusing conjecture. But for \)d=4,5\( we require local stationarity as well as the vanishing of additional derivatives as in \eqref{eq:highD}, as well as a dominant energy condition \eqref{gDEC}. Under the above conditions, we also showed the universal sector of leading-order holographic theories to satisfy a finite renormalized QNEC. The success of this derivation for \)d\le 3\( (using only \)\theta=0\() suggests that the QNEC may hold for general field theories in contexts where it cannot be derived from the quantum focusing conjecture (QFC). If so, it would be incorrect to think of the QFC as being more fundamental than the QNEC; the QNEC seems to have a life of its own. For \)d\ge6\( we argued these properties to generally fail even for weakly isolated horizons (where all derivatives of \)\theta, \sigma_{ab}\( vanish) satisfying the dominant energy condition, though they do hold on Killing horizons. The issue in \(d=6\) is that finite counter-terms in the effective action can contain derivatives of the Riemann tensor, and that these terms change the definition of the entropy and stress tensor in such a way that the combination entering the QNEC is not invariant. Our \(d=5\) argument for scheme-independence required the conditions [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"} and [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"}, while for \(d=4\) we need only [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"} and the first line of [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"}. It certainly appears that local stationarity is not itself sufficient for the four-derivative counter-terms in [\[eq:d45\]](#eq:d45){reference-type="eqref" reference="eq:d45"}, this has been shown (using and in footnote [\[Fufoot\]](#Fufoot){reference-type="ref" reference="Fufoot"}) only for \(d\ge 5\) in the case of the Gauss-Bonnet term. For \(d=4\) the Gauss-Bonnet term gives \(\Delta Q=0\). Terms involving only the scalar curvature (i.e., the \(R^2\) term) are easily handled by changing conformal frame to write the theory as Einstein-Hilbert gravity coupled to a scalar field. So if the QNEC is invariant under the remaining \(R_{ab}R^{ab}\) term, one would expect a useful QNEC (and thus perhaps also a useful quantum focussing condition (QFC)) to hold in \(d=4\) as well. As explained in, the QNEC implies the perturbative semi-classical generalized second law (GSL) of thermodynamics at first non-trivial order in the gravitational coupling \(G\). A consequence of our work is a thus proof of the (first-order) semi-classical GSL on causal horizons satisfying the conditions above, and in particular on general causal horizons for \(d\le 3\). Even at this order, this is the first GSL proof valid when the null congruence \(N\) does not reduce to a Killing horizon in the background. Now, as described in footnote [\[Fufoot\]](#Fufoot){reference-type="ref" reference="Fufoot"}, it was recently shown in that for \(d \ge 5\) the QNEC generally fails to be scheme-independent when the change of renormalization scheme entails the addition of a Gauss-Bonnet term to the action, and furthermore that the associated change \(\Delta Q\) can have either sign. It then follows that (for theories that require a Gauss-Bonnet counter-term) a renormalized QNEC cannot hold in general renormalization schemes as, if one finds a finite \(Q \ge 0\) with some scheme, we may always change the scheme to add a Gauss-Bonnet term so that \(Q_{\rm modified} = Q + \Delta Q_{\rm GB} < 0\). In a rather different direction, it was recently noted that on Killing horizons the quantum null energy condition is related to a property of the relative entropy \(S(\rho||\sigma)\) between an arbitrary state \(\rho\) and the vacuum state \(\sigma\): \[\begin{aligned}
0 \leq T_{kk}-\frac{\hbar}{2\pi} S'' = S(\rho||\sigma)'' \end{aligned}\] In this equation, the derivatives of the relative entropy are the same type of local derivatives with respect to null deformations of the region that appear in \(S''\). (This "concavity" property is about the *second* derivative, while the well-known monotonicity of relative entropy bounds the *first* derivative, \(S(\rho||\sigma)' \leq 0\).) While we have seen that the QNEC is not always well defined or true in curved space, the relative entropy is known to be scheme-independent. It would thus be interesting to understand if an inequality of the form \(S(\rho||\sigma)'' \geq 0\) for appropriate \(\sigma\) might hold more generally, perhaps even in cases where the QNEC fails. One might also investigate whether, without introducing any smearing, this could lead to a new conjecture for theories with dynamical gravity that could replace the quantum focussing condition and which might hold even when the original QFC is violated. Finally, we comment on the relation of the QNEC to the observation of that at least the QFC violation of can be avoided by taking the QFC to apply only to suitably smooth variations of the generalized entropy. One might then ask if such a "smeared QFC" could lead to a suitably smeared version of the QNEC that would hold even when the original QNEC fails. However, the original QFC reduces to the QNEC only at locally stationary points where \(\theta |_p= \dot{\theta}|_p= \sigma_{ab}|_p=0\). And the point of the averaging in is precisely that, when \(d > 3\) and \(k^a h_e^{~b}h_f^{~c}h_g^{~d}R_{abcd} \neq 0\) (where \(k^a\) and \(h_{a}{}^{b}\) are respectively the null normal vector and the projector onto the chosen cut of the null congruence \(N\)), the locally stationary condition can hold only a set of measure zero. Dimensional analysis then shows that smearing out the QFC on scales long compared to the cutoff leads manifestly non-positive QFC contributions from violations of local stationarity to swamp those from failures of the QNEC. In other words, any QNEC-like inequality is irrelevant to the smeared QFC of unless \[\label{QNECcond} d\le 3\text{ or }k^a h_e^{~b}h_f^{~c}h_g^{~d}R_{abcd}=0,\] so that only under one of these conditions could any QNEC be derived from this smeared QFC. We therefore suspect that these are the most general conditions under which any QNEC could possibly hold, and the analysis of section [2.2.3](#scheme-indep){reference-type="ref" reference="scheme-indep"} suggests that further conditions are likely required at least for \(d\ge 6\). Indeed, as shown in appendix [\[appendix:identities\]](#appendix:identities){reference-type="ref" reference="appendix:identities"}, the condition [\[QNECcond\]](#QNECcond){reference-type="eqref" reference="QNECcond"} follows from [\[eq:highD\]](#eq:highD){reference-type="eqref" reference="eq:highD"}, and [\[gDEC\]](#gDEC){reference-type="eqref" reference="gDEC"}, and then from one sees that it suffices to avoid the \(d\ge 5\) QNEC violation associated with the Gauss-Bonnet term. However, it remains to further investigate the effect of the \(R_{ab}R^{ab}\) counter-term for both \(d=4,5\) because [\[QNECcond\]](#QNECcond){reference-type="eqref" reference="QNECcond"} does not appear to guarantee the existence of a null congruence \(N\) satisfying the sufficient conditions which we used in this paper. Similar comments must also apply to the proposed "quantum dominant energy condition" of, which reduces to the QNEC when considering pairs of variations that act in the same null direction. | {'timestamp': '2017-12-29T02:02:17', 'yymm': '1706', 'arxiv_id': '1706.01572', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01572'} |
# Introduction
Nowadays deep laser cooling of neutral atoms is routinely used for broad range of modern quantum physics researches including metrology, atom optics, and quantum degeneracy studies. The well-known techniques for laser cooling below the Doppler limit, like sub-Doppler polarization gradient cooling , velocity selective coherent population trapping or Raman cooling are restricted to atoms with degenerated over angular momentum energy levels or hyperfine structure. However, for atoms with single ground state \(^{24}\)Mg, \(^{40}\)Ca, \(^{88}\)Sr, \(^{174}\)Yb are of interest for developing optical time standard these techniques can not be applied directly. For example, for \(^{24}\)Mg atoms with the ground state \(^1S_0\) the Doppler cooling temperature (\(k_B T_D \approx \hbar \gamma/2\)) can be reached on closed singlet transition \(^1S_0 \rightarrow \, ^{1}P_1\) (\(\lambda = 285.3\) nm). For lower temperature additional cooling on \(^3P_2 \rightarrow \, ^{3}D_3\) optical transition with degenerated over angular momentum energy levels can be applied. However, the experimental realization of laser cooling on \(^3P_2 \rightarrow \, ^{3}D_3\) optical transition does not result significant progress. The atoms were cooled to temperature \(T \approx 1\, mK\) is about Doppler limit only. The quantum simulation of laser cooling are also shows the limitation of cooling temperature to about Doppler limit in conventional MOT, formed by laser waves with circular polarization. An alternative way of deep laser cooling of these elements is to use narrow lines and "quenching" techniques of narrow-line laser cooling successfully applied for \(^{40}\)Ca atoms but, to our knowledge, still do not show significant progress for \(^{24}\)Mg atoms. Recently, laser-driven Sisyphus-cooling scheme was proposed for cooling atoms in optical dipole trap. This scheme utilize the difference in trap-induced ac Stark shift for ground and exited levels of atom coupled by resonant laser light. The laser cooling scheme has clear semiclassical interpretation: been excited by resonant laser light on the bottom of shallow optical potential related to the ground state an atom moves further in steepest potential related to excited state. Spontaneous emission returns it back to the shallow potential in the ground state. The loosing a portion of energy in each act of this process results to atom cooling after several cycles due to "Sisyphus effect". This semiclassical model was applied for description of laser cooling of Yb and Sr in optical dipole trap. In the following paper we study application of this cooling scheme to \(^{24}\)Mg atom on narrow \((3s3s)\,^1S_0 \rightarrow \, (3s3p)\,^{3}P_1\) (\(\lambda_1 = 457\) nm, \(\gamma_1 = 196\) s\(^{-1}\)) optical transition. Additional light field resonant to \((3s3p)\,^3P_1 \rightarrow \, (3s4s)\,^{1}S_0\) optical transition (\(\lambda_2 = 462\) nm, \(\gamma_{21} = 109\) s\(^{-1}\) and \(\gamma_{23} = 2.1\cdot 10^7\) s\(^{-1}\)) is applied for optical quenching (see Fig.[\[mg_levels\]](#mg_levels){reference-type="ref" reference="mg_levels"}), i.e. increasing the effective linewidth of optical transition. We find the semiclassical description of laser cooling of Mg atom with narrow optical transition can't be used here. For description of laser cooling we use quantum approaches that allow to take into account optical pumping and photon recoil effects in laser cooling process. In the paper we point our attention to minimum laser cooling temperature for described scheme and cooling time as well.
# Description of the model
We consider the motion of \(^{24}\)Mg atom in the dipole optical trap with \(\lambda_{D} = 1064 \,nm\) that provide higher polarizability of atom in the excited state \((3s3p)\,^{3}P_1\) than in the ground state \((3s3s)\,^1S_0\). In the following paper we restrict our consideration by two-level model assuming the quench field results to increasing effective linewidth of optical transition to \(\gamma_{eff}\) : \[\gamma_{eff} = \gamma_1 + \gamma_2 \frac{\Omega_2^2}{\gamma_2^2+4\, \delta_2^2}\, ,\] where \(\Omega_2\) is Rabi and \(\delta_2\) is detuning of quench field. Thus, for example, to get \(\gamma_{eff} = 100 \gamma_1\) at \(\delta_2 = 0\) one have to apply the quench field intensity \(I_{q} \approx 1.6 \,W/cm^2\). The simulated polarizability difference for the optical dipole trap wavelength is about \(\alpha_e/\alpha_g = 1.46\). In the trap (fig.[\[egen_states\]](#egen_states){reference-type="ref" reference="egen_states"}) the quantum nature of atomic motion becomes essential. For considered optical trap depth of the ground state \(U_g = 200\, \mu K\) with vibration energy separation of the lowest states are \(\hbar \omega_g \approx 16.6\, \mu K\). The excited state optical potential depth \(U_e = 292\, \mu K\) and the lowest states separation are \(\hbar \omega_e \approx 20.1 \,\mu K\). The large energy levels separation require quantum model for description of laser cooling dynamics of atoms in the trap. Really, the semiclassical models can't be applied here because of the velocity range of the semiclassical damping force on Mg atoms has sharp variation in momentum space \(\Delta p \ll \hbar k\) (see figure [\[mg_force\]](#mg_force){reference-type="ref" reference="mg_force"}). As well the semiclassical parameter \(\varepsilon_R = \hbar k^2 /(2M\gamma) \gg 1\) is not small, that also contradicts requirements for semiclassical approach. For description of laser cooling of Mg in the optical trap we consider two quantum approaches. The first one is based on decomposition of atom density matrix on optical potential vibration level states. Restricting by limited number of lowest vibration states we simulate the stationary distribution over the vibration levels in the trap, as well as the laser cooling dynamics to steady state distribution. This approach is similar to method was described in. However, in our model we also take into account the optical coherence of different vibration states. The second method we consider is based on direct numerical solution of quantum equation for atom density matrix that allows to take into account not only the fixed number of the lowest vibration level states but whole density matrix of atoms, which also include tunneling effects and above barrier motion. However, in this method, due to the high complicity of the problem we omit the recoil effects from the pumping field that is equivalent to orthogonal orientation of wave vectors of pumping and optical trap light waves in one dimensional model.
## Two-level model: exact numerical solution of quantum density matrix equation
We consider the motion of Mg atom in the optical dipole trap is standing light wave propagating along \(z\) direction with linear polarization along \(x\). The pumping light field also linear polarized along \(x\) with wavevector along \(z\) or \(y\). The quantum equation for atom density matrix describes evolution of internal and external states of atoms \[\label{QQE} \frac{\partial}{\partial t}{\hat \rho} =-\frac{i}{\hbar} \left[{\hat H}_0 + {\hat V}_{ed},{\hat \rho}\right] +{\hat \Gamma}\left\{{\hat \rho}\right\}\] with \({\hat H}_0\) is Hamiltonian, \({\hat V}_{ed}\) describes interaction with pumping field and \({\hat \Gamma}\left\{{\hat \rho}\right\}\) describes relaxation of density matrix due to spontaneous decay. As was mentioned above, we restrict our consideration by effective two-level model with \((3s3s)\,^1S_0\) is the ground (g) and \((3s3p)\,^{3}P_1\) is excited state (e), assuming the influence of the quench field \(\Omega_2\) results to adjustable linewidth by modification of decay rate from \(\gamma_1\) to \(\gamma_{eff}\) only, as described in. Further in the paper we omit parameters indexes \(\Omega_1\), \(\delta_1\) and \(\gamma_1\) by writing \(\Omega\), \(\delta\) and \(\gamma\) instead. The Hamiltonian of atom in the trap has the form: \[{\hat H}_0 = \frac{{\hat p}^2}{2M} +\hbar \omega_g(z) |g\rangle \langle g | +\hbar [\omega_e(z) +\omega_0]|e\rangle \langle e |\] with optical potentials in the ground \(\hbar \omega_g(z) = U_g \cos^2(kz)\) and \(\hbar \omega_e(z) = U_e \cos^2(kz)\) in excited states (\(\hbar \omega_0 = E_e-E_g\) is energy difference of unperturbed ground and excited states). The wavevector \(k = 2\pi/\lambda_{D}\) is defined by the dipole trap. Applying rotating wave approximation the equation for atom density matrix components in coordinate representation \({\hat \rho}(z_1,z_2)\) takes the followign form: \[\begin{aligned}
\label{basic_eq} \left(\frac{\partial}{\partial t}-\frac{i\hbar}{M} \frac{\partial}{\partial q} \frac{\partial}{\partial z}\right) \rho^{ee} &=&-\gamma_{eff}\, \rho^{ee} +i\frac{U_e}{\hbar} \sin(2kz)\sin(kq) \rho^{ee}-\frac{i}{\hbar} \left[ {\hat V} \rho^{ge}-\rho^{eg} {\hat V}^{\dagger} \right] \nonumber \\ \left(\frac{\partial}{\partial t}-\frac{i\hbar}{M} \frac{\partial}{\partial q} \frac{\partial}{\partial z}\right) \rho^{gg} &=& {\hat \gamma}\{ \rho^{ee} \} +i\frac{U_g}{\hbar} \sin(2kz)\sin(kq) \rho^{gg}-\frac{i}{\hbar} \left[ {\hat V}^{\dagger} \rho^{eg}-\rho^{ge} {\hat V} \right] \nonumber \\ \left(\frac{\partial}{\partial t}-\frac{i\hbar}{M} \frac{\partial}{\partial q} \frac{\partial}{\partial z}\right) \rho^{eg} &+&\left(\frac{\gamma_{eff}}{2}-i\tilde{\delta}(z,q)\right)\rho^{eg} =-\frac{i}{\hbar} \left[ {\hat V} \rho^{gg}-\rho^{ee} {\hat V} \right] \nonumber \\ \left(\frac{\partial}{\partial t}-\frac{i\hbar}{M} \frac{\partial}{\partial q} \frac{\partial}{\partial z}\right) \rho^{ee} &+&\left(\frac{\gamma_{eff}}{2}+i\tilde{\delta}(z,-q)\right)\rho^{ge} =-\frac{i}{\hbar} \left[ {\hat V}^{\dagger} \rho^{ee}-\rho^{gg} {\hat V}^{\dagger} \right]\,, \end{aligned}\] with \(z = (z_1+z_2)/2\), \(q = z_1-z_2\), and the function \(\tilde{\delta}(z,q)\): \[\tilde{\delta}(z,q) = \delta-(U_e-U_g)\frac{1+\cos(2kz)\cos(kq)}{2\hbar} +(U_e+U_g)\frac{\sin(2kz)\sin(kq)}{2\hbar}.\] The spontaneous income part to the ground state \({\hat \gamma}\) in coordinate representation for two-level model has simple form : \[\begin{aligned}
\label{sp_income} {\hat \gamma}\{\rho^{ee}\} &=& {\tilde{\gamma}}(q) \rho^{ee} \nonumber \\ {\tilde{\gamma}}(q) &=& 3\,\gamma_{eff}\left( \frac{\sin(k_1q)}{(k_1q)^3}-\frac{\cos(k_1q)}{(k_1q)^2} \right) \end{aligned}\] with \(k_1 = 2\pi/\lambda\) is wavevector of emitted photon. The pumping field induces transitions between the ground and excited states. This part is described in ([\[basic_eq\]](#basic_eq){reference-type="ref" reference="basic_eq"}) by operator \[\label{Vpump} {\hat V} = \Omega/2\, \exp(ik_1 z \cos(\theta))\] with \(\Omega\) is Rabi frequency of pumping field and \(\theta\) is angle between the axis \(z\) and pumping wave propagation direction. For the case of orthogonal orientation of the pumping wave propagation to the dipole trap \(\theta = \pi/2\) the equation for density matrix ([\[basic_eq\]](#basic_eq){reference-type="ref" reference="basic_eq"}) can be solved numerically by the method suggested in. It should be noted the considered method allows to get steady state solution for density matrix with taking into account quantum recoil effects as for atoms in the trap as for nontrapped atoms. The figure [\[mg_dist\]](#mg_dist){reference-type="ref" reference="mg_dist"} shows spatial and momentum distribution of Mg atoms in the optical dipole trap for orthogonal orientation of pumping wave (\(\theta = \pi/2\)), for pumping field intensity \(I \simeq 34\) \(W/cm^2\) (\(\Omega/\gamma = 20000\)) and different detunings. The obtained numerical solution for steady state density matrix \({\hat \rho}(z_1,z_2)\) contains whole information on internal and external states of atoms in the trap. In particular one can extract the population of vibration levels in the ground and excited states: \[\begin{aligned}
\rho^{ee}_n &=& \int \psi^{*(e)}_n(z_1)\rho^{ee}(z_1,z_2)\psi^{(e)}_n(z_2) dz_1 dz_2 \, , \nonumber \\ \rho^{gg}_n &=& \int \psi^{*(g)}_n(z_1)\rho^{gg}(z_1,z_2)\psi^{(g)}_n(z_2) dz_1 dz_2 \end{aligned}\] where \(\psi^{(e,g)}_n(z)\) are n-th vibration level eigenfunctions. The distribution of vibration levels population in the ground and excited states for parameters of figure [\[mg_dist\]](#mg_dist){reference-type="ref" reference="mg_dist"} are shown on figure [\[mg_popdist\]](#mg_popdist){reference-type="ref" reference="mg_popdist"}. The energy of cooled atoms can be found by different way. First of all one can use the relation for the temperature of cooled atoms in the well known form \[\label{Eclass0} k_B T = <p^2>/M\,,\] with \(<p^2> = Tr\{{\hat p}^2\, {\hat \rho}\}\). This relation neglects the atom localization effects in the optical potential. The most accurate relation for energy is expressed by the following averaging: \[\label{Eclass} E = Tr\left\{ \left( \frac{{\hat p}^2}{2M}\, + \hbar \omega_e(z)\right){\hat \rho^{(ee)}}\right\} +Tr\left\{ \left( \frac{{\hat p}^2}{2M}\, + \hbar \omega_g(z)\right){\hat \rho^{(gg)}}\right\}\.\] As an alternative way one can find the average energy over the vibration states \[\label{E_vibr} E = \sum_n E^{(e)}_n \, \rho^{(ee)}_n +E^{(g)}_n \, \rho^{(gg)}_n.\] For considered parameters all above definitions give very close values that denotes the main contribution to energy are given by atoms on the lowest vibration energy levels in the region where the optical potential has close to parabola shape. The energy of Mg atoms for different detuning is shown on figure [\[model1_dd\]](#model1_dd){reference-type="ref" reference="model1_dd"}. The total population of excited vibration level states here do not exceed \(2\%\) and are not shown on figure [\[model1_dd\]](#model1_dd){reference-type="ref" reference="model1_dd"}(b). In the region of detuning \(\delta >-6000 \gamma\) we find inversion of the lowest vibration levels population resulting to energy growth. For the higher intensity of pumping field this effect is also exists and moves to larger detuning area. The energy of atoms as function of pumping field intensity for different detunings is shown on figure ([\[model1_Om\]](#model1_Om){reference-type="ref" reference="model1_Om"}).
## Decomposition on vibration states model
As we see from the simulations above based on numerical solution of basic equation for atomic density matrix ([\[basic_eq\]](#basic_eq){reference-type="ref" reference="basic_eq"}) for the considered parameters Mg atoms can be cooled and well localized in the dipole trap. Thus for description of laser coolling and laser cooling time we can also apply an alternative approach based on decomposition of atom density matrix over vibration level states. \[{\hat \rho} = \left( \begin{array}{cc} \left \begin{array}{ccc} {\rho}^{ee}_{00} & {\rho}^{ee}_{01} &... \\ {\rho}^{ee}_{10} & {\rho}^{ee}_{11} &... \\ ... &... & {\rho}^{ee}_{nm} \\ \end{array}\right] & \left \begin{array}{ccc} {\rho}^{eg}_{00} & {\rho}^{eg}_{01} &... \\ {\rho}^{eg}_{10} & {\rho}^{eg}_{11} &... \\ ... &... & {\rho}^{eg}_{nm} \\ \end{array}\right] \\ \left \begin{array}{ccc} {\rho}^{ge}_{00} & {\rho}^{ge}_{01} &... \\ {\rho}^{ge}_{10} & {\rho}^{ge}_{11} &... \\ ... &... & {\rho}^{ge}_{nm} \\ \end{array}\right] & \left \begin{array}{ccc} {\rho}^{gg}_{00} & {\rho}^{gg}_{01} &... \\ {\rho}^{gg}_{10} & {\rho}^{gg}_{11} &... \\ ... &... & {\rho}^{gg}_{nm} \\ \end{array}\right] \\ \end{array \right)\] The equation for these components takes the form: \[\label{vibr_eq} \frac{\partial }{\partial t}{\hat \rho} =-\frac{i}{\hbar}\left[{\hat H}_0+{\hat W},{\hat \rho} \right] +\Gamma\{{\hat \rho}\}\] with Hamiltonian \[{\hat H}_0 = \left( \begin{array}{cc} \left \begin{array}{ccc} E^{(e)}_{0}-\hbar\delta & 0 &... \\ 0 & E^{(e)}_{1}-\hbar\delta &... \\ ... &... & E^{(e)}_{n}-\hbar\delta \\ \end{array}\right] & \left \begin{array}{ccc} 0 & 0 &... \\ 0 & 0 &... \\ ... &... & 0 \\ \end{array}\right] \\ \left \begin{array}{ccc} 0 & 0 &... \\ 0 & 0 &... \\ ... &... & 0 \\ \end{array}\right] & \left \begin{array}{ccc} E^{(g)}_0 & 0 &... \\ 0 & E^{(g)}_1 &... \\ ... &... & E^{(g)}_n \\ \end{array}\right] \\ \end{array \right)\] where \(E^{(e)}_n\) and \(E^{(g)}_n\) are the vibration levels energy of the excited and the ground states. The pumping light field to atom interaction part has nondiagonal block elements only: \[W^{eg}_{nm} = \langle e,n| {\hat V} |g,m \rangle = \int \psi^{*(e)}_n(z) {\hat V} \psi^{(g)}_m(z) dz \,.\] with \({\hat V}\) is defined in ([\[Vpump\]](#Vpump){reference-type="ref" reference="Vpump"}). The spontaneous relaxation part \(\Gamma\{{\hat \rho}\}\) has a standard form \[\Gamma\{ {\hat \rho} \} =-\frac{\gamma_{eff}}{2} \left\{ |e,n \rangle \langle e,n|,{\hat \rho} \right\} + {\hat \gamma} \left\{ {\hat \rho} \right\}\] with \({\hat \gamma} \left\{ {\hat \rho} \right\}\) describes income to the ground vibration states has the following matrix elements: \[\begin{aligned}
{\hat \gamma} \left\{ {\hat \rho} \right\}_{nm} &=& \sum_{\nu \mu} \Gamma^{\nu \mu}_{nm} {\rho}^{ee}_{\nu \mu} \nonumber \\ \Gamma^{\nu \mu}_{nm} & = & \int \psi^{*(g)}_n(z_1)\psi^{(e)}_{\nu}(z_1) {\tilde{\gamma}}(z_1-z_2) \psi^{*(e)}_n(z_2)\psi^{(g)}_{\nu}(z_2) dz_1 dz_2 \end{aligned}\] and \({\tilde{\gamma}}(q)\) is defined in ([\[sp_income\]](#sp_income){reference-type="ref" reference="sp_income"}). Thus in the equation ([\[vibr_eq\]](#vibr_eq){reference-type="ref" reference="vibr_eq"}) we take into account the evolution of diagonal elements of density matrix (vibration levels population) and nondiagonal elements as well. However, compare to exact numerical solution described above we should restrict our consideration by limited number of vibration levels neglecting tunneling effects and above barrier motion. Here bellow we consider 10 vibration levels on the ground and excited states. The atom steady state energy as function of pumping field intensity is shown on figure [\[model2_Om\]](#model2_Om){reference-type="ref" reference="model2_Om"}(a). Even the model is simplified it gives the cooling energy result close to the direct numerical solution of eq.([\[basic_eq\]](#basic_eq){reference-type="ref" reference="basic_eq"}) (see figure [\[model1_Om\]](#model1_Om){reference-type="ref" reference="model1_Om"}) at \(\theta =\pi/2\). The difference appears at large pumping field intensity where the populations of the top vibration levels are not negligible and tunneling effects can not be neglected, i.e. far from the field parameters required for cooling to minimum energy. Additionally this model allows to solve dynamical problem and estimate the cooling time. To find the cooling time we assume the atoms populate the highest vibrational energy level of the ground state optical potential at \(t = 0\). The time evolution of vibration levels population has a complex dependence. We fit \({\rho}^{gg}_{00}(t)\) by exponential function of the form \({\rho}^{gg}_{00}(t) = a-b \, \mbox{exp}(-t/{\tau})\) with \(\tau\) describes the cooling time. Additionally we note, the energy of cooled atoms does not depend on parameter \(\gamma_{eff}\) (i.e. on quench field intensity) in the range of our simulations \(\gamma < \gamma_{ef} < 100 \gamma\), while the cooling time \(\tau\) is inversely proportional to \(\gamma_{eff}\) in considered model. This allows us to represent the cooling time \(\tau\) in the more general form through dimensionless value \(\tilde{\tau}\) figure [\[model2_Om\]](#model2_Om){reference-type="ref" reference="model2_Om"}(b) \[\tau = \tilde{\tau} /\gamma_{eff}.\] The cooling time and the energy of atoms as function of two parameters: pumping field detuning and intensity are shown on figure [\[results_2dOm\]](#results_2dOm){reference-type="ref" reference="results_2dOm"}. The conditions for minimum cooling energy do not coincide with conditions for fast cooling. For orientation angle \(\theta = 0\) the minimum of cooling energy \(E_{min} = 8.4 \mu K\) is reached for pumping field intensity \(I \simeq 0.1 W/cm^2\) and detuning \(\delta/\gamma \simeq-8050\), while the minimum cooling time \(\tau\gamma_{eff} \simeq 7.4\) is reached at \(I \simeq 28.5 W/cm^2\) and detuning \(\delta/\gamma \simeq-2987\) (figure [\[results_2dOm\]](#results_2dOm){reference-type="ref" reference="results_2dOm"}(a,b)). For orthogonal orientation \(\theta = \pi/2\) the minimum energy \(E_{min} = 13.2 \mu K\) is reached for \(I \simeq 0.1 W/cm^2\) and detuning \(\delta/\gamma \simeq-17500\), but the cooling time for these parameters is extremely large \(\tau\gamma_{eff} \simeq 2 \cdot 10^5\). For field parameters providing a reasonable cooling time \(\tau\gamma_{eff} <100\) the energy above \(60 \mu K\) can be reached only (figure [\[results_2dOm\]](#results_2dOm){reference-type="ref" reference="results_2dOm"}(c,d)).
# Conclusion
We study laser cooling of \(^{24}Mg\) atoms in the dipole trap with pumping field resonant to narrow \((3s3s)\,^1S_0 \rightarrow \, (3s3p)\,^{3}P_1\) (\(\lambda = 457\) nm) optical transition and quench field resonant to \((3s3p)\,^3P_1 \rightarrow \, (3s4s)\,^{1}S_0\). The effect of quenching we consider as widening of optical transition to \(\gamma_{eff}\) only. We find the semiclassical model can not be used for description of laser cooling in this scheme. We suggest quantum models. The first one is based on the direct numerical solution of quantum kinetic equation for atom density matrix and the second is simplified model is based on decomposition of atom density matrix over vibration states in the dipole trap. The second model has limitations and describes the cooling of atoms on the lowest vibration levels only. The results of this model is differ from exact numerical solution for enough high intensity of pumping field (above 50\(W/cm^2\) or Rabi above \(25000 \gamma\)) when populations of the top vibration levels are not negligible, i.e. tunneling effects and above barrier motion of atoms can not be neglected. Nevertheless the simplified model well describes the laser cooling of atoms cooled to minimum energy. Additionally, the simplified model allows to estimate the cooling time. The parameters of pumping field for cooling to minimum energy do not coincide with conditions for fast cooling. We find parameters that allow cooling the atoms at reasonable cooling time \(\tau \simeq 10 / \gamma_{eff}\) (i.e. \(\tau \simeq 0.5\, ms\) at \(\gamma_{eff} = 100\gamma\)) to energy \(E\simeq 12 \, \mu K\). In the considered models the steady state solution do not depends on \(\gamma_{eff}\) while the cooling time is inversely proportional to \(\gamma_{eff}\). The difference may appears for more complex model that takes into account pumping to level \((3s4s)\,^{1}S_0\) by quench field. We consider this question in the next paper. The work was supported by Russian Science Foundation (project N 16-12-00054). V.I.Yudin acknowledges the support of the Ministry of Education and Science (3.1326.2017). | {'timestamp': '2017-06-07T02:03:06', 'yymm': '1706', 'arxiv_id': '1706.01593', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01593'} |
null | null |
Subsets and Splits